# How Do You Build a Useful Robot?
Canonical source: [https://isaiuseful.com/robotics](https://isaiuseful.com/robotics)
Robotics builder’s guide + field watch · checked 1 August 2026
The useful robotics startup is not “a humanoid for everything.” It is one expensive, repetitive or unsafe task—inside one measurable environment—with a recovery path and economics that survive contact with the real world.
- [Choose a first task](#route)
- [Browse useful robots](#mainstream-robots)
- [See the humanoid frontier](#frontier)
- [**≈90K ft²** Apptronik Robot Park VENDOR-REPORTED TRAINING FACILITY](https://apptronik.com/news-collection/welcome-to-robot-park-where-apptroniks-apollo-goes-to-work)
- [**100K+ totes** Digit at GXO VENDOR-REPORTED LIVE-DEPLOYMENT OUTPUT](https://www.agilityrobotics.com/content/digit-moves-over-100k-totes)
- [**13,361** UWORLD U1 orders UBTECH-REPORTED AT LAUNCH · NOT DELIVERIES](https://www.prnewswire.com/news-releases/ubtech-launches-uworld-u1-the-worlds-first-full-size-mass-produced-ultra-bionic-humanoid-robot-302815285.html)
- [**10 h shifts** Figure 02 at BMW CUSTOMER-REPORTED PILOT SCOPE](https://www.bmwgroup.com/en/news/general/2026/humanoid-robot-in-leipzig.html)
- [**1M+ traces** AgiBot World OPEN ROBOT-LEARNING DATASET](https://github.com/OpenDriveLab/Agibot-World)
- [**12 releases** Figure · 2023 → 2026 BODY → VLA → BOTQ → FIELD WORK](https://www.figure.ai/news/ramping-figure-03-production)
Humanoids · August 2026
## What makes a humanoid robot useful at work?
Humanoid teams are coupling dexterous hardware, vision-language-action models, whole-body control, large robot datasets and production engineering. The real dividing line is no longer a polished motion—it is sustained work with a denominator.
- [Video: Brett Adcock on Figure’s eight-hour autonomous shift](https://www.youtube.com/watch?v=og0VF9-XCSs)
Figure package-sorting endurance test ·
14 May 2026
### From an eight-hour demo to named customer work.
Watch the endurance run here; use the named BMW and Catalyst deployments below to judge whether the capability is becoming sustained customer work.
What is independently attributable
The interview follows Figure’s autonomous package-sorting endurance run. Figure did not identify that May test as work for a postal carrier or at a customer site. BMW separately says Figure 02 worked ten-hour weekday shifts during a ten-month production pilot, moving more than 90,000 parts in about 1,250 operating hours.
On 26 May, Catalyst Brands announced a commercial Figure deployment at its Reno distribution center for sorting and packing around its Joey Pouch system—the closest verified customer match to the package-sorting demo.
**Evidence boundary:** Neither Figure nor Catalyst says the livestreamed endurance test was commissioned by Catalyst. The agreement also does not disclose robot count, intervention rate, uptime or all-in cost, so the announcement and demo remain separate evidence.
- [Read BMW’s pilot account →](https://www.bmwgroup.com/en/news/general/2026/humanoid-robot-in-leipzig.html)
- [Read Catalyst’s customer announcement →](https://corporate.jcpenney.com/2026/05/26/catalyst-brands-taps-figure-ai-for-humanoid-automation/)
- [Inspect Figure’s package-sorting note →](https://www.figure.ai/news/helix-logistics)
01
**Whole-body autonomy**
Coordinate feet, torso, arms, hands and balance from pixels instead of stitching together isolated skills.
02
**Dexterous sensing**
Add touch, force and palm-level vision for contact-rich work, occlusion and deformable objects.
03
**Data engines**
Scale teleoperation, human-motion transfer, simulation, failure capture and reusable robot datasets.
04
**Long-horizon work**
Preserve task state, recover from variation and complete minutes or shifts—not edited single attempts.
05
**Industrialization**
Drive cost, reliability, charging, serviceability, safety integration and manufacturing volume together.
**Do not stop at the video.**
1. Demo What capability is being claimed?
2. Technical note What data, architecture and test were used?
3. Open artifact Can code, weights, data or a benchmark be inspected?
4. Customer account Does the operator confirm the task and duration?
5. Field denominator Success, interventions, uptime, safety and cost per accepted cycle.
Figure · release timeline
## Three years of iteration, in order.
The sequence matters more than any single clip: body, factory hardware, upper-body VLA, data scaling, deformable objects, home hardware, full-body autonomy and a return to production. Every result below is vendor evidence unless a customer source is explicitly linked.
**2023**
- [Video: Introducing Figure](https://www.youtube.com/watch?v=b37rQZ4maPo)
01 Mar 2023
Mission
### Introducing Figure
The starting thesis: put a commercially useful general-purpose humanoid into human environments.
- [Figure release archive →](https://www.figure.ai/news)
**2024**
- [Video: Introducing Figure 02](https://www.youtube.com/watch?v=0SRVJaOg9Co)
06 Aug 2024
Hardware
### Introducing Figure 02
A second-generation body aimed at the factory: actuation, hands, sensing, compute and cable routing move toward deployment.
- [Read the later BMW results →](https://www.figure.ai/news/production-at-bmw)
**2025**
- [Video: Introducing Helix](https://www.youtube.com/watch?v=Z3yQHYNXPws)
20 Feb 2025
Generalist control
### Introducing Helix
One VLA connects language and vision to 200 Hz continuous upper-body control; Figure reports training on roughly 500 hours of teleoperated data.
- [Read the technical note →](https://www.figure.ai/news/helix)
- [Video: Scaling Helix Logistics](https://www.youtube.com/watch?v=lkc2y0yb89U)
07 Jun 2025
Data scaling
### Scaling Helix · logistics
More demonstrations plus visual memory, state history and force feedback target package variety, speed and barcode orientation.
- [Inspect results + ablations →](https://www.figure.ai/news/scaling-helix-logistics)
- [Video: Scaling Helix Laundry](https://www.youtube.com/watch?v=HOoRnv3lA0k)
12 Aug 2025
Deformables
### Scaling Helix · laundry
The same architecture, with a new dataset, moves from rigid parcels to cloth that folds, slips and changes geometry.
- [Read the method note →](https://www.figure.ai/news/helix-learns-to-fold-laundry)
- [Video: Scaling Helix Dishes](https://www.youtube.com/watch?v=8gfuUzDn4Q8)
03 Sep 2025
Household task
### Scaling Helix · dishes
Bimanual grasping, object placement and recovery move the learning system into fragile, cluttered household work.
- [Read the evaluation note →](https://www.figure.ai/news/helix-loads-the-dishwasher)
- [Video: Introducing Figure 03](https://www.youtube.com/watch?v=Eu5mYMavctM)
09 Oct 2025
Co-designed body
### Introducing Figure 03
Palm cameras, tactile fingertips, softer coverings, wireless charging and manufacturing design bring the body closer to Helix and the home.
- [Read the hardware account →](https://www.figure.ai/news/introducing-figure-03)
**2026**
- [Video: Introducing Helix 02](https://www.youtube.com/watch?v=lQsvTrRTBRs)
27 Jan 2026
Full body
### Introducing Helix 02
Figure reports a unified pixels-to-actions system controlling locomotion, manipulation and balance, including a four-minute 61-action kitchen sequence.
- [Read architecture + tests →](https://www.figure.ai/news/helix-02)
- [Video: Helix 02 Living Room Tidy](https://www.youtube.com/watch?v=CAdTjePDBfc)
09 Mar 2026
Long-horizon home
### Helix 02 · living room tidy
Room-scale navigation and object return test whether the full-body policy can preserve task state across a changing scene.
- [Read Figure’s test account →](https://www.figure.ai/news/helix-02-living-room-tidy)
- [Video: BotQ Ramping Figure 03 Production](https://www.youtube.com/watch?v=YZH1csMhnDo)
29 Apr 2026
Production system
### BotQ ramps Figure 03
Figure reports more than 350 third-generation robots delivered from BotQ and a demonstrated line cadence increase from one robot per day to one per hour. That is factory throughput—not proof of autonomous customer utilization.
- [Inspect yield, testing + fleet claims →](https://www.figure.ai/news/ramping-figure-03-production)
- [Video: Helix 02 Bedroom Tidy](https://www.youtube.com/watch?v=8xEuFQz4E4A)
08 May 2026
Generalization
### Helix 02 · bedroom tidy
A second room and new object distribution probe whether the same learned system transfers beyond one staged environment.
- [Read Figure’s test account →](https://www.figure.ai/news/helix-02-bedroom-tidy)
- [Video: Figure 03 at BMW](https://www.youtube.com/watch?v=tv90GFM9RAo)
30 Jun 2026
Return to factory
### F.03 at BMW
Helix 02 and Figure 03 meet a sequencing workflow that combines thin-part picking, stepping, placement and cart manipulation.
- [Read Figure’s account →](https://www.figure.ai/news/f-03-at-bmw)
- [Check BMW’s media record →](https://www.press.bmwgroup.com/global/video/detail/PF0010224/Figure-03-at-BMW-Group-Plant-Spartanburg)
Humanoid builders · different bets
## Home, factory, cost, data—and talent.
These companies are not racing on one scoreboard. The humanoid field spans supervised home service, vertically integrated manufacturing, cheaper hardware, commercial logistics, robot-data factories, consumer companions and the builder networks that supply the next generation of talent.
- [Video: Introducing Calvin-40](https://www.youtube.com/watch?v=YddS-aI097Q)
Wandercraft · Calvin‑40
### Turn medical-grade balance and torque into industrial work.
Founded in 2012, Wandercraft built hands-free, self-balancing medical exoskeletons before adding an upper body for Calvin. The company says that platform is engineered to move users weighing up to 100 kg; its safety-critical history is relevant provenance, not proof that Calvin is the industry’s most stable humanoid. Renault reports testing Calvin at Douai on a roughly 30 kg tire-handling task and targets about ten robots by the end of 2026 and 350 across French and Spanish factories by the end of 2027. Those are customer roadmap figures, not completed deployment, reliability or task economics.
- [Inspect the exoskeleton lineage →](https://en.wandercraft.eu/)
- [Inspect the current Calvin product route →](https://en.wandercraft.eu/calvin-40)
- [Read Renault’s factory account →](https://www.renaultgroup.com/en/magazine/technology/calvin-a-new-generation-robot-is-born/)
- [Video: NEURA Gym: The First Physical AI Training Center for Robots](https://www.youtube.com/watch?v=hUNujYlRmZU)
NEURA Robotics · NEURA Gym + 4NE1
### Make real-world training infrastructure part of the product.
NEURA announced a Series C with a total round size of up to $1.4 billion in June 2026; that is a maximum round size, not a $1.4 billion lifetime total or proof that every dollar has already translated into deployed robots. The supplied film presents NEURA Gym as a place for robots to collect multimodal real-world data and for companies to train applications. NEURA’s current route describes partner-specific training, validation and integration facilities opening across Europe and the United States. The infrastructure and funding are substantial, but neither supplies 4NE1 uptime, intervention rate or customer throughput.
- [Inspect the current NEURA Gym route →](https://neura-robotics.com/neuragym/)
- [Read the exact Series C disclosure →](https://neura-robotics.com/record-series-c/)
- [Video: UMA unveils its vision for the next generation of robots, Remi Cadene at Machina Summit](https://www.youtube.com/watch?v=0vc-dAQDIQc)
UMA · Northstar + real-time learning
### Combine open robot-learning talent with a new body.
UMA’s founding team includes LeRobot builders from Hugging Face and a founding member of Google DeepMind’s robotics team. The supplied July 2026 talk presents its Northstar design and a learning approach intended to acquire skills from human demonstration rather than task-by-task programming. UMA’s current site says logistics and manufacturing pilot programs will launch during 2026. A prototype and planned pilots are not product shipments: no public unit count, customer operating record, uptime or task economics is yet available.
- [Verify the team + pilot status →](https://uma.bot/)
- [Video: GENE.01 Design Concept at CES 2026](https://www.youtube.com/watch?v=zYRz1uIIpwQ)
Generative Bionics · GENE.01
### Make touch part of the body—then prove the industrial task.
The supplied CES film shows the January 2026 design concept; Generative Bionics now presents GENE.01 as a sensorized platform with tactile, force and vision sensing and publishes an open URDF model. Fincantieri independently confirms a four-year program to adapt the platform for shipyard welding, with first on-site tests scheduled by the end of 2026. That is credible current development and an inspectable artifact—not certification, deployment uptime or welding performance.
- [Inspect the current platform + robot model →](https://gbionics.ai/gene01/)
- [Read Fincantieri’s validation plan →](https://www.fincantieri.com/en/newsroom/press-releases/2026/fincantieri-and-generative-bionics-launch-an-industrial-partnership-to-develop-a-humanoid-welding-robot-for-shipyards)
- [Video: PAL Robotics Kangaroo Pro](https://www.youtube.com/watch?v=G7riD4yb1tg)
PAL Robotics · KANGAROO Pro
### Expose a configurable humanoid research platform.
Barcelona-based PAL Robotics still markets KANGAROO Pro and lists ROS 2, MuJoCo, Gazebo and Isaac Lab support across Lite, Standard and Plus configurations. The platform is positioned for research and advanced manipulation rather than verified production work; the published specifications and film do not supply uptime, intervention or customer-throughput denominators.
- [Inspect the current product →](https://pal-robotics.com/robot/kangaroo/)
- [Read the current datasheet →](https://pal-robotics.com/wp-content/uploads/2025/05/2025-Datasheet_ENG_KANGAROO-PRO.pdf)
- [Video: Meet Genesis AI Eno](https://www.youtube.com/watch?v=zab62_u-a5U)
- [Video: Announcing Genesis World 1.0](https://www.youtube.com/watch?v=ggfsGD1PCQA)
Genesis AI · Eno + Genesis World 1.0
### Co-design the body, hand, model and evaluation engine.
Eno uses a wheeled, height-adjustable body that folds compactly and carries proprietary human-scale hands with 20 active, back-drivable degrees of freedom. Genesis plans targeted customer deployments by the end of 2026, so this is a product roadmap rather than field proof. Its open Genesis World 1.0 stack makes a complementary bet: use GPU simulation for closed-loop evaluation, while training on real data. Genesis reports reducing a typical 200-plus-hour hardware evaluation to under 0.5 simulated hours; the company’s own sim-to-real calibration and task suite still set that comparison.
- [Read Eno’s launch + deployment plan →](https://www.genesis.ai/press/meet-eno)
- [Inspect Genesis World’s method →](https://www.genesis.ai/blog/the-role-of-simulation-in-scalable-robotics-genesis-world-10-and-the-path-forward)
- [Open the simulator repository →](https://github.com/Genesis-Embodied-AI/genesis-world)
- [Video: ▶ Europe’s robotics clubs](https://www.youtube.com/watch?v=OHCf75KS8i4)
Europe · builder network
### The frontier is also who gets a robot into their hands.
Prototype’s March 2026 account describes the European Student Robotics Association as 13 clubs across eight countries with more than 2,500 students. That is an ecosystem claim, not a product benchmark, but it matters: cheaper platforms, shared labs and cross-border teams determine how quickly Europe converts research talent into repeatable hardware practice.
- [Read the sourced club list →](https://updates.prototypecap.com/p/the-kids-are-building)
- [Open ESRA →](https://www.studentrobotics.eu/)
- [Video: Physical AI in Action at NVIDIA GTC 2026 | Humanoid Robotics Demo](https://www.youtube.com/watch?v=0oZAw6rryIE)
Humanoid · HMND 01 + KinetIQ
### Scale the company and fleet brain before commercial rollout.
Founded in 2024, Humanoid disclosed a $152 million Series A in July 2026, bringing its company-reported total raised to $270 million. The supplied GTC film shows two wheeled, gripper-equipped robots in a simulated store responding to voice requests, allocating work and coordinating handoffs through KinetIQ. That is a live multi-robot demonstration, not evidence of bipedal field autonomy. Humanoid says Beta robot rollout begins in Q4 2026, so the funding and industrial agreements are capacity signals rather than delivered-unit, uptime or cost-per-task evidence.
- [Read the GTC demonstration account →](https://thehumanoid.ai/humanoid-brings-voice-controlled-robotic-fulfillment-powered-by-kinetiqto-life-at-nvidia-gtc/)
- [Check funding + rollout timing →](https://thehumanoid.ai/humanoid-raises-152-million-at-1-35-billion-post-money-valuation-becoming-europes-first-pure-play-humanoid-robotics-unicorn/)
- [Video: Reflect v1.0 - The Path Towards Long-Horizon Autonomous Humanoid Work](https://www.youtube.com/watch?v=6dme3JYj3Hg)
Flexion · Reflect v1.0
### Compose navigation, manipulation and recovery into one mission.
The supplied film puts one humanoid through a multi-floor parcel mission involving stairs, an elevator, doors, tool use and recovery from errors after a single instruction. Flexion reports 90% end-to-end completion with reinforcement learning versus 38% with supervised fine-tuning alone on one 16-step internal evaluation. Its technical note also names the limits: a bounded task distribution, difficult grasps, visual assumptions and incomplete recovery coverage. This is a useful integrated vendor test of the intelligence stack—not independent replication or a customer production denominator.
- [Inspect architecture, evaluation + limits →](https://flexion.ai/news/flexion-reflect-v1.0)
- [Video: WSJ tries NEO](https://www.youtube.com/watch?v=f3c4mQty_so)
- [Video: 1X launch film](https://www.youtube.com/watch?v=LTYMWadOW7c)
1X · NEO
### Ship the home service before full autonomy.
1X’s own product page says early owners receive basic autonomy while unfamiliar complex tasks can be handled through scheduled remote expert supervision. That is a real product architecture—not a footnote. Redwood is described as a 160M-parameter onboard policy running at roughly 5 Hz; privacy, teleoperation availability and the rate at which remote help declines are central performance metrics.
- [Inspect NEO + Expert Mode →](https://www.1x.tech/neo)
- [Read the Redwood technical note →](https://www.1x.tech/discover/redwood-ai)
- [Video: Tesla Bot update](https://www.youtube.com/watch?v=XiQkeWOFwmk)
- [Video: Optimus Gen 2](https://www.youtube.com/watch?v=cpraXaw7dyc)
Tesla · Optimus
### Reuse the real-world AI and manufacturing machine.
The strategic bet is vertical: vision-based autonomy, in-house compute, actuators, factories and manufacturing scale under one roof. But the supplied public videos are from 2023. Tesla’s 2025 annual filing still describes Optimus as a humanoid “in development,” while its 2025 year-end update lists California Optimus manufacturing capacity as under construction. Treat production timing and capability forecasts as targets until deployed task data appears.
- [Read Tesla’s 2025 annual filing →](https://ir.tesla.com/_flysystem/s3/sec/000162828026003952/tsla-20251231-gen.pdf)
- [Check the capacity disclosure →](https://ir.tesla.com/_flysystem/s3/sec/000162828026003837/tsla-20260128-gen.pdf)
- [Video: Form and function of enterprise humanoid design: Boston Dynamics Atlas](https://www.youtube.com/watch?v=wL0-Pu_8F0U)
Boston Dynamics · Atlas
### Design the humanoid around industrial work.
Spun out of MIT’s Leg Lab by Marc Raibert in 1992, Boston Dynamics is based in Waltham, Massachusetts. Hyundai Motor Group owns 80% after a 2021 transaction that valued the company at $1.1 billion; SoftBank retains 20%. Production Atlas is 1.9 m tall, weighs 90 kg and lists 56 degrees of freedom, a four-hour battery and 30 kg sustained payload. Boston Dynamics says its 2026 production is committed to Hyundai’s RMAC and Google DeepMind. These are company specifications and scheduled deployments—not customer-reported uptime, intervention or cost-per-task.
- [Inspect Atlas specifications →](https://bostondynamics.com/products/atlas/)
- [Read the design brief →](https://bostondynamics.com/webinars/form-function-enterprise-humanoid-design/)
- [Review the company history →](https://bostondynamics.com/about/)
- [Check the deployment status →](https://bostondynamics.com/blog/boston-dynamics-unveils-new-atlas-robot-to-revolutionize-industry/)
- [Video: Gemini Robotics 2 brings whole body intelligence to robots](https://www.youtube.com/watch?v=4lSQnrMC6nY)
Apptronik · Apollo 2 + Robot Park
### Build a data factory around the fleet.
Apptronik’s expanded Austin Robot Park is a nearly 90,000 ft² training facility for fleets of Apollo 2 robots in bipedal and wheeled configurations. The company says related workflows run at Google DeepMind and customer sites including Mercedes‑Benz and GXO, mixing teleoperation, autonomous execution and simulation to train Gemini Robotics. Mercedes‑Benz independently confirms using Apollo in production-data collection and intralogistics tests. Neither source publishes fleet size, autonomy share, intervention rate or customer throughput.
- [Read Apptronik’s Robot Park report →](https://apptronik.com/news-collection/welcome-to-robot-park-where-apptroniks-apollo-goes-to-work)
- [Inspect the Apollo 2 platform →](https://apptronik.com/apollo/apollo-2)
- [Review Gemini Robotics 2’s results →](https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/)
- [Check Mercedes‑Benz’s account →](https://group.mercedes-benz.com/company/production/procuction-network/mbdfc-humanoid-robots.html)
- [Video: What Will You Build With Us?](https://www.youtube.com/watch?v=LM3hTXUShvA)
Agility Robotics · Digit
### Turn one logistics task into a commercial service.
Digit’s evidence is unusually operational for a humanoid: Agility reports more than 100,000 totes moved in GXO’s live facility, and Toyota Motor Manufacturing Canada signed a Robots‑as‑a‑Service agreement in February 2026 after a pilot. The Toyota agreement confirms a commercial step but omits robot count, price and deployment-wide reliability. The tote milestone is also vendor-reported and does not disclose the time period, intervention rate or number of robots behind the total.
- [Inspect the 100,000-tote claim →](https://www.agilityrobotics.com/content/digit-moves-over-100k-totes)
- [Read the Toyota agreement →](https://www.agilityrobotics.com/content/agility-robotics-announces-commercial-agreement-with-toyota-motor-manufacturing-canada)
- [Video: Introducing Generalist GEN-1](https://www.youtube.com/watch?v=SY2xyrmV44Y)
Generalist · GEN‑1
### Pretrain on human interaction, then adapt to the robot.
Generalist says GEN‑1 was pretrained on more than 500,000 hours of human physical-interaction data with no robot data, then adapted to each shown task with about one hour of robot data. Its own six-task evaluation reports 99% average success versus 64% for the previous GEN‑0 and roughly three-times-faster completion on selected tasks. The company also names unsolved tasks and offers early access; the figures remain vendor-run, task-specific evaluations rather than customer production metrics.
- [Inspect runs, baselines + limitations →](https://generalistai.com/blog/gen-1)
- [Video: Unitree R1](https://www.youtube.com/watch?v=mTMYfVD4zCw)
- [Video: Unitree GD01](https://www.youtube.com/watch?v=oWOyUMJWptc)
Unitree · R1 + GD01
### Compress the hardware cost—and widen the form factor.
R1 starts at a vendor-listed $4,900 before tax and shipping, weighs roughly 27–29 kg and lists about one hour of battery life. At the other extreme, GD01 is a 500 kg-class rideable biped/quadruped machine. Together they show Unitree’s hardware breadth, not general autonomy. Unitree itself warns that humanoids remain early-stage and that some demonstrated functions are still under development.
- [Check R1 configurations + caveats →](https://www.unitree.com/R1/)
- [Review Unitree’s release record →](https://www.unitree.com/about/)
- [Video: First batch of Henan-made EngineAI T800 robots rolls off the line](https://www.youtube.com/watch?v=1bTC_qfF7sA)
EngineAI · T800
### Pair athletic hardware with a manufacturing ramp.
Established in October 2023, headquartered in Shenzhen and led by founder and CEO Zhao Tongyang, EngineAI reports RMB 1 billion in cumulative Pre‑A++ and A1 funding, followed by A1+ and A2 rounds whose amounts it did not disclose. It lists four T800 editions from $40,500. The official video shows a first Henan-made batch leaving the Zhengzhou line; separately, the company says its 12,000 m² Shenzhen base can complete one humanoid every 15 minutes and supports a 10,000-unit scalable-delivery capability. Those are vendor capacity claims—not shipment totals, utilization data or customer-verified autonomous work.
- [Check T800 editions + caveats →](https://en.engineai.com.cn/product-t800.html)
- [Review the company profile →](https://en.engineai.com.cn/about-us)
- [Inspect the funding disclosure →](https://en.engineai.com.cn/about-news-media/38.html)
- [Inspect the capacity claim →](https://www.prnewswire.com/news-releases/engineai-launches-shenzhen-intelligent-manufacturing-base-as-first-batch-of-t800-humanoid-robots-roll-off-the-production-line-to-begin-mass-delivery-302785446.html)
- [Video: ▶ AGIBOT in precision manufacturing](https://www.youtube.com/watch?v=h6rCRa8qUFw)
AGIBOT · deployment + open data
### Scale the data engine and the factory rollout together.
AGIBOT’s strongest signal is the pairing of artifacts and operations. AgiBot World exposes more than one million robot trajectories across 217 tasks and five deployment scenarios. Separately, the company reports about 100 cumulative hours of a G2 factory livestream and a 15,000th robot production milestone in June 2026. Production count is not autonomous-task success, so keep shipment, utilization and intervention metrics separate.
- [Read AGIBOT’s production report →](https://www.agibot.com/article/231/detail/82.html)
- [Inspect data + GO-1 code →](https://github.com/OpenDriveLab/Agibot-World)
- [Read the AgiBot World paper →](https://arxiv.org/abs/2503.06669)
- [Video: Meet LimX Luna](https://www.youtube.com/watch?v=-lgo5xqgVko)
LimX Dynamics · Luna
### Package motion intelligence for public spaces.
Luna is a 160 cm, 56 kg interactive humanoid aimed at malls, museums, theme parks and live stages—not a factory generalist. LimX lists 27 active body degrees of freedom, up to 3 kg arm payload and about four hours of battery life, measured in its lab. The standard configuration uses fist-shaped ends; a six-DoF five-finger hand is optional. Video imitation, a no-code task editor and 200-plus-unit synchronized control are vendor capabilities, not independently measured autonomy or uptime.
- [Read the Luna launch →](https://www.limxdynamics.com/en/news/BK000062)
- [Check configurations + lab specs →](https://www.limxdynamics.com/en/products/luna/spec)
- [Video: Why XPENG IRON is so human-like](https://www.youtube.com/watch?v=ssxdSVQN14I)
XPENG · next-generation IRON
### Reuse an EV company’s chips, AI and production discipline.
XPENG’s 2025 model is specified at 82 whole-body degrees of freedom, including 22 per hand, with three in-house Turing AI chips. The company says IRON is being tested in its own production environment and targets mass production by the end of 2026. That target remains prospective: public material does not give delivered units, sustained autonomous cycles, intervention rate or all-in task cost.
- [Inspect XPENG’s technical claims →](https://www.xiaopeng.com/news/company_news/5511.html)
- [Read XPENG’s 2025 ESG disclosure →](https://ir.xiaopeng.com/static-files/10e70f1b-cbed-4fb1-9f2f-f6be521d72a2)
- [Video: Astribot Lumo-2](https://www.youtube.com/watch?v=Q0WMUDF3JKI)
Astribot · Lumo-2
### Reason over future dynamics without rendering the future.
Founded in Shenzhen in December 2022, Astribot pairs its S1 manipulation platform with Lumo‑2, a latent world-action model introduced in a July 2026 paper. The authors report gains over named VLA and world-action baselines on long-horizon and dexterous real-world tasks by aligning actions with learned latent dynamics, vision and language. This is author-reported research evidence; the release does not establish customer throughput, deployment uptime or broadly available model weights.
- [Read the Lumo‑2 paper →](https://arxiv.org/abs/2607.11270)
- [Review Astribot’s company account →](https://www.astribot.com/en/about)
- [Video: UBTECH Launches UWORLD U1](https://www.youtube.com/watch?v=jXRbNaFqByo)
UBTECH · UWORLD U1
### Move the humanoid pitch from industry to companionship.
UBTECH’s June 2026 launch introduces three U1 models spanning a semi-torso Lite and two full-body editions, with vendor pricing from RMB 119,800. The company reports 88 degrees of freedom and 13,361 cumulative orders at launch. Those are demand and specification claims—not delivered units, independent capability tests or evidence that the “mass-produced” positioning has translated into sustained work. Consumer interaction also creates a different evidence burden around privacy, reliability and safety.
- [Inspect UBTECH’s launch claims →](https://www.prnewswire.com/news-releases/ubtech-launches-uworld-u1-the-worlds-first-full-size-mass-produced-ultra-bionic-humanoid-robot-302815285.html)
**Evidence policy:** company pages and product films establish what each vendor claims; customer accounts from BMW, Mercedes‑Benz, GXO, Renault, Fincantieri and Toyota’s quoted manufacturing leadership corroborate specific pilots, development programs or agreements. Filings constrain Tesla and XPENG’s public status; Humanoid and NEURA’s funding disclosures, Flexion’s internal evaluation, UMA’s pilot timing, Apptronik’s fleet size, Agility’s tote denominator, UBTECH’s orders, EngineAI’s capacity, Figure’s production ramp and the new model benchmarks remain vendor claims or incompletely specified roadmaps. Open repositories and papers make AgiBot, Astribot, GENE.01 and Genesis World artifacts more inspectable. None substitutes for a site-specific safety case or your own uptime and cost measurements.
Mainstream robotics · narrow jobs
## Most useful robots do not look human.
Surgical systems, warehouse fleets, service carts, robot arms, construction printers and autonomous mowers trade generality for a bounded job. “Mainstream” describes this mixed field—not equal maturity, autonomy or return on investment.
- [Video: GEBHARDT InstaPick warehouse robotics](https://www.youtube.com/watch?v=NuI9ACH0Gjk)
GEBHARDT · InstaPick · commercial
### Move standard bins without building a humanoid.
GEBHARDT continues to market InstaPick as a modular robot-based container storage system and integrates it with picking and conveyor technology. The meaningful comparison is against alternative storage and material-flow designs: capacity, fire strategy, replenishment, failure isolation, service access and peak throughput.
- [Inspect current warehouse robotics →](https://gebhardt-group.com/index.php/en/products/warehouse-technology-systems/warehouse-robotics.html)
- [Video: Timo Boll versus a KUKA robot](https://www.youtube.com/watch?v=lv6op2HHIuM)
KUKA · industrial automation · established
### A spectacular film is not the industrial use case.
The table-tennis film demonstrates motion and marketing, not a production benchmark. KUKA is an active global automation group with industrial robots, mobile robots, software and integration services; evaluate the exact cell, payload, reach, safety design, cycle time, tooling and integrator support instead of generalizing from the stunt.
- [Verify the current company →](https://www.kuka.com/en-us/company)
- [Browse the current portfolio →](https://www.kuka.com/en-us)
- [Video: Brightpick robotic warehouse](https://www.youtube.com/watch?v=U2AGLeJBFNg)
Brightpick · fulfilment robots · commercial
### Bring picking to the aisle instead of people to the goods.
Brightpick remains active and launched Gridpicker in 2026 after Autopicker and Giraffe. Its films and customer announcements show a coherent commercial system; the buyer still needs SKU-level success, human touches, exception labor, storage density, peak throughput and service response measured on the proposed design.
- [Inspect current systems →](https://brightpick.ai/)
- [Check current customer + product news →](https://brightpick.ai/news-releases/)
- [Video: Ocado Smart Platform warehouse automation](https://www.youtube.com/watch?v=Eu3jgy2-tL8)
Ocado Group · fulfilment platform · deployed
### Make the fleet, grid, arms and software one operating system.
Ocado still sells the Smart Platform and now also describes a mobile-robot system with Chuck and Porter. The film is useful precisely because no single robot explains the outcome: storage geometry, orchestration, picking, packing, software and site operations determine throughput together.
- [Inspect the current robotics stack →](https://www.ocadogroup.com/about-us/our-technology)
- [Review the commercial platform →](https://www.ocadogroup.com/our-solutions/online-grocery)
- [Video: Robotic hernia repair surgery](https://www.youtube.com/watch?v=KAvQsRL-jeo)
Intuitive · da Vinci · commercial
### A mature robot can be highly specialized and human-controlled.
Sutter Health’s film shows robotic-assisted hernia repair; the da Vinci system is made by Intuitive, which continues to manufacture, place and support surgical systems in 2026. This is not autonomous surgery: a trained surgeon controls the instruments, and clinical procurement requires procedure-specific evidence, training, cybersecurity and service planning.
- [Inspect the current systems →](https://www.intuitive.com/)
- [Check the 2026 filing →](https://isrg.intuitive.com/static-files/8d536618-2150-41fb-b463-a6d6691d59bd)
- [Video: Introducing the ICON Titan Print System](https://www.youtube.com/watch?v=Ebir6dl_cxc)
ICON · Titan · commercial rollout
### Turn 3D printing into a supported construction system.
ICON commercially launched Titan in March 2026 with reservations, training planned for Q3 2026 and first deliveries anticipated in early 2027. The vendor lists the printer, pump, mixer, software, materials and support as one program; cost and speed claims still need validation against local design, permitting, foundations, reinforcement, finishes and crew requirements.
- [Verify rollout + delivery timing →](https://www.iconbuild.com/newsroom/icon-announces-first-commercial-rollout-of-its-3d-printing-construction-technology-for-builders)
- [Inspect the current program →](https://www.iconbuild.com/technology/program)
- [Video: AmbiSort AI-powered robotic parcel sorting](https://www.youtube.com/watch?v=iYkg3woXB1A)
Ambi Robotics · AmbiSort · commercial
### Automate one parcel touch inside an existing operation.
AmbiSort combines a robot arm, vision, grasp planning, barcode scanning and configurable destinations for small-parcel sortation. Ambi Robotics still markets the A‑Series and announced a 2026 integration with Pickle Robot; throughput gains shown by the vendor remain configuration- and parcel-mix-dependent.
- [Inspect the current A‑Series →](https://www.ambirobotics.com/ambisort-a-series/)
- [Check current deployments + partnerships →](https://www.ambirobotics.com/media-room/press/)
- [Video: Introducing Almond Axol](https://www.youtube.com/watch?v=WjcBwSufr7Q)
Almond · Axol · early commercial
### Sell a dual-arm data and manipulation platform.
Y Combinator lists Almond as active, and Almond currently offers Axol with an open SDK, published configurations, pricing and short stated lead time. It is a young four-person company, so treat availability, support, repair, payload and integration claims as procurement questions—not the same maturity class as an established industrial vendor.
- [Inspect Axol + current ordering →](https://www.almond.bot/)
- [Verify the active YC profile →](https://www.ycombinator.com/companies/almond-2)
- [Video: Scythe autonomous commercial mower](https://www.youtube.com/watch?v=dysTBGW6g_A)
Scythe · M.52 · acquired, supported
### Acquisition did not end the product.
ASI acquired Scythe Robotics in March 2026, and the company says M.52 remains an operating equipment brand with continuing customer support, field service and development. This is landscaping rather than crop agriculture; assess terrain, boundary handling, transport, charging, blade service, public exposure and operator supervision.
- [Verify the acquisition + support path →](https://scytherobotics.com/blog-posts/scythe-acquired-by-asi)
- [Video: KEENON XMAN-R1 robot barista](https://www.youtube.com/watch?v=oTZFSlYEBSk)
- [Video: KEENON DINERBOT T9](https://www.youtube.com/watch?v=9FQzaSydT-c)
KEENON · service robots · marketed
### Separate beverage preparation from table delivery.
KEENON currently markets XMAN‑R1 alongside DINERBOT, BUTLERBOT, KLEENBOT and industrial-delivery products. The two films show different automation layers: a robot barista and an indoor delivery platform. Buyer evidence should cover site mapping, queue behavior, cleaning, refill labor, food safety and human recovery—not only the serving motion.
- [Review KEENON’s current range →](https://www.keenon.com/en/product/XMAN-F1/)
- [Video: ROKAE robotic welding](https://www.youtube.com/watch?v=DlHE3zcauBk)
ROKAE · welding robots · active
### Package vision and force control around a skilled trade.
ROKAE still sells industrial and collaborative welding systems and published updated 2026 manuals, certifications and product material. The vendor’s trade-show demonstrations establish current activity, not weld qualification on your parts; test seam extraction, fixturing, consumables, spatter, rework and operator training.
- [Inspect the current welding line →](https://www.rokae.com/en/product/show/726.html?ly=serch)
- [Check the 2026 company update →](https://www.rokae.com/en/news/show/2458/ROKAE-Robotics-at-ITES-Shenzhen-2026.html)
- [Video: Meet Beni by Mondo Robotics](https://www.youtube.com/watch?v=rBETwu7ssVw)
Mondo Robotics · Beni · crowdfunding
### A live company is not yet a mainstream product.
Mondo Robotics is active and reports more than 30 beta users, but Beni launched through Kickstarter in July 2026. The camera rover belongs on the watchlist with a clear crowdfunding label: independent durability, privacy, repair, software-support and delivery evidence are still thin.
- [Check the company timeline →](https://mondorobotics.com/pages/about-us)
- [Review current updates →](https://mondorobotics.com/blogs/news)
- [Video: temi personal robot](https://www.youtube.com/watch?v=jNriaEZkNJI)
temi · mobile telepresence · active
### Mobility, video and reminders can be enough.
temi remains an active self-navigating service and telepresence platform, with current deployments in care, education and commercial settings. It has no arms and should not be sold as a general household worker; evaluate navigation coverage, privacy, remote administration, fall or obstruction handling, and the staff workflow around it.
- [Review the current company →](https://www.robotemi.com/about/)
- [Inspect the product routes →](https://www.robotemi.com/)
**Current-only policy:** every included company has a current official product, support, filing, acquirer or accelerator record checked on 28 July 2026. A compelling film without a verified current company and product route is omitted.
Agriculture robotics · field + greenhouse
## Useful farm robots are crop- and task-specific.
Weeding, harvesting, planting, spraying and pest control fail in different ways. Every card below represents a current company and a current product or active development program.
- [Video: Agrobot robotic strawberry harvester](https://www.youtube.com/watch?v=M3SGScaShhw)
Agrobot · strawberry harvesting · development
### Give each berry its own perception-and-cut problem.
Agrobot remains active in Spain and California and describes a newer solar-powered, modular-arm harvester. Its public material does not establish broad commercial availability or customer throughput, so the machine stays in the development lane until field yield, damage, miss rate, speed and service data are available.
- [Review the current company account →](https://www.agrobot.com/copia-de-about-us)
- [Video: AgXeed AgBot field use case](https://www.youtube.com/watch?v=ATqTBUedyjM)
AgXeed · AgBot · commercial
### Automate the tractor route and keep standard implements.
AgXeed actively markets four AgBot models, TraXwise planning software and a dealer network, with 2026 demonstrations and product news. Vendor savings and utilization claims require farm-specific validation across implements, soil, weather, refueling, road transport, connectivity, supervision and local machinery rules.
- [Inspect the current AgBot range →](https://www.agxeed.com/)
- [Check 2026 activity →](https://www.agxeed.com/news-and-events/)
- [Video: VitiBot Bakus vineyard robot](https://www.youtube.com/watch?v=v5oig0P9X2Y)
VitiBot · Bakus · SDF group
### Design the vehicle around vineyard rows.
VitiBot remains active inside SDF Group and markets narrow- and wide-vineyard Bakus models with dealer and demonstration routes. Its electric autonomous straddle form is specific to vineyard geometry; validate slopes, row width, tools, battery duty, safety zoning and dealer support locally.
- [Inspect the current Bakus range →](https://vitibot.fr/)
- [Video: PATS-X bat-like pest-control drones](https://www.youtube.com/watch?v=4g5Sb-UVn6I)
PATS · pest-control drones · final testing
### Target airborne pests inside the greenhouse.
PATS remains active, but PATS‑X is still described as undergoing final testing in Dutch and Belgian crops and is offered via a waitlist. The film is a development demonstration; effectiveness, non-target impact, crop coverage, maintenance and operating supervision need field evidence.
- [Check current test + waitlist status →](https://www.pats-drones.com/pats-x)
- [Review the company →](https://www.pats-drones.com/pats)
- [Video: Ridder and MetoMotion GRoW tomato harvesting robot](https://www.youtube.com/watch?v=-FB7mjc9eQ8)
Ridder + MetoMotion · GRoW · pre-order
### Pick, collect and box vine tomatoes in one pass.
Ridder and MetoMotion remain active, and Ridder currently accepts reservations for GRoW. The two-arm robot has named greenhouse trials and a customer scale-up account, but published labor and cost reductions remain vendor estimates; require current cycle, damage, intervention, variety and service data.
- [Check current reservation status →](https://info.ridder.com/reserve-your-grow-tomato-harvesting-robot)
- [Inspect the current system →](https://metomotion.com/robotic-worker/)
- [Video: TTA-ISO CuttingPlanter 2.0](https://www.youtube.com/watch?v=QcqKduExS0U)
TTA‑ISO · nursery automation · commercial
### Automate cuttings without pretending the whole greenhouse is autonomous.
TTA‑ISO is active and introduced CuttingPlanter 2.0 in July 2026, adding vision software, a faster arm and an electric tilting gripper. It sells a broader nursery-automation portfolio; compare plant material, tray formats, changeover, rejects, sanitation and upstream/downstream labor.
- [Verify the July 2026 launch →](https://tta-iso.com/updates/introducing-our-newest-cuttingplanter-20)
- [Browse the current equipment range →](https://www.tta-iso.com/)
- [Video: FarmBot open-source farming robot](https://www.youtube.com/watch?v=yHy8rJpbxH4)
FarmBot · garden CNC · commercial + open source
### Make small-plot automation inspectable and hackable.
FarmBot continues to sell hardware and update its app and operating system in 2026. Its CAD, software and developer routes are open, which improves inspectability and education; it does not turn a garden-scale gantry into a field-scale commercial farming system.
- [Inspect software + open-source routes →](https://farm.bot/pages/software)
- [Check current product updates →](https://farm.bot/blogs/news)
- [Video: AGRIST cucumber harvesting robot](https://www.youtube.com/watch?v=bylYcrRl9ik)
AGRIST · greenhouse harvesting · active
### Combine harvesting with routine crop care.
AGRIST is active and in May 2026 announced a funded program to extend its cucumber robot from harvesting into leaf removal and fruit thinning. That is forward development, not proof the three-task system is already deployed; ask for crop variety, greenhouse geometry, success rate, cycle time and human reset data.
- [Check the May 2026 program →](https://agrist.com/archives/14976)
**Current-only policy:** every entry must have a current company plus a current product or active development program. Superseded machines, acquired IP, research-only prototypes, category mismatches and unverifiable businesses stay off the public catalogue.
Food + drink robotics
## The hard part is the whole kitchen shift.
Cooking motion is only one layer. Ingredients, cold chain, sanitation, allergen controls, replenishment, waste, recipes, service, approvals and human recovery decide whether the system is useful.
- [Video: goodBytz robotic kitchen](https://www.youtube.com/watch?v=GiG6Kmz_FfE)
goodBytz · autonomous kitchen · commercial
### Use deployments, not awards, as the stronger signal.
goodBytz is active and announced a July 2026 U.S. defence-site delivery after deployments in Europe and South Korea. Its system cooks, portions and serves; site buyers still need accepted-meal throughput, ingredient labor, cleaning, allergens, downtime, local approvals and service data.
- [Check current products + deployments →](https://www.goodbytz.com/)
- [Inspect the current press record →](https://www.goodbytz.com/de/press)
- [Video: Eatch robotic kitchen for large-scale cooking](https://www.youtube.com/watch?v=qtkONuUDwHI)
Eatch · production kitchen · active
### Scale individually cooked meals from a central kitchen.
Eatch remains active and markets its Robotic Kitchen Technology for central production and white-label meals. Its claim is flexible meal production rather than a public-facing robot restaurant; evaluate recipe range, batch planning, input prep, packaging, sanitation, maintenance and delivered-food quality.
- [Inspect the current platform →](https://eatch.me/)
- [See the operating food route →](https://maaltijden.eatch.me/pages/onze-robot)
- [Video: Moley Robotics showroom](https://www.youtube.com/watch?v=WzeGT3m5oMU)
Moley Robotics · luxury kitchen · marketed
### A showroom product needs a home-service model.
Moley remains active, showed the system at Salone del Mobile in 2026 and markets a built-in robotic kitchen. This is a high-end fitted installation, not a countertop appliance; procurement hinges on kitchen integration, recipe authoring, cleaning, utensils, child safety, remote support and long-term parts availability.
- [Check current company activity →](https://www.moley.com/news/)
- [Inspect the current kitchen →](https://www.moley.com/)
- [Video: Richtech Robotics ADAM at the Vegas Golden Knights](https://www.youtube.com/watch?v=cPXQFNWisAQ)
Richtech Robotics · ADAM · public company
### Make the robot both service equipment and a visible attraction.
Richtech is an active Nasdaq filer and continues to promote ADAM in 2026. The arena film proves a branded event installation, not beverage-unit economics; verify drink quality, queue time, consumables, cleaning, staffing, uptime and the difference between entertainment value and operational savings.
- [Check current company releases →](https://ir.richtechrobotics.com/news-events/news-releases)
- [Read the 2026 annual filing →](https://ir.richtechrobotics.com/static-files/778957e1-6ab8-40e9-a349-a16482278166)
- [Video: BOTINKIT OMNI cooking robot](https://www.youtube.com/watch?v=Kcw0kw2v8ps)
BOTINKIT · OMNI · commercial
### Standardize wok cooking while people run the kitchen.
BOTINKIT is active, has a Japanese subsidiary and says OMNI is deployed across Asian and North American markets. The automated cooking station is one component of a kitchen; test recipes, ingredient prep, chef controls, extraction, cleaning, certification, spare parts and regional service.
- [Verify company + market presence →](https://www.botinkit.co.jp/about)
- [Inspect the current platform →](https://www.botinkit.ai/)
**Current-only policy:** every entry must have a current operating company and a current supported product. Liquidated operations, acquired technology without a current product route and stale operating evidence stay off the public catalogue; duplicate films are consolidated.
Market scale · counts with caveats
## A small builder field can serve a very large machine economy.
There is no audited global census of “robotics companies,” and databases classify vendors, products and integrators differently. The order of magnitude is still revealing: specialist supplier landscapes number in the hundreds, while deployed machines already number in the millions.
**≈700**
warehouse-automation ecosystem companies
LogisticsIQ’s landscape, cited by ITIF in 2023, is broad: it includes robots, software, components, integrators, infrastructure and related services—not 700 interchangeable robot startups.
- [Inspect the category definition →](https://itif.org/publications/2023/06/26/policymakers-should-support-robotic-automation-to-solve-productivity-crunch-in-logistics/)
**≈200**
companies developing humanoids
A 2026 industry account cites Gartner’s “nearly 200” estimate. Gartner’s own outlook is sharper: by 2028, it expects fewer than 100 companies to move proofs of concept beyond experiments and fewer than 20 to reach production in manufacturing or supply chain.
- [Trace the ≈200 estimate →](https://www.machinedesign.com/markets/robotics/article/55363632/physical-ai-hype-vs-reality-kung-fu-robots-are-coolbut-should-you-hire-one)
- [Read Gartner’s production forecast →](https://www.gartner.com/en/newsroom/press-releases/2026-01-21-gartner-predicts-fewer-than-20-companies-will-scale-humanoid-robots-for-manufacturing-and-supply-chain-to-production-stage-by-2028)
**15,384**
commercial martech solutions in 2025
This is a product catalog, not a company count, so it is not an apples-to-apples market share calculation. It is a useful competition benchmark: one mature software function supports tens of thousands of discoverable products.
- [Open the 2025 landscape method →](https://chiefmartec.com/2025/05/2025-marketing-technology-landscape-supergraphic-100x-growth-since-2011-but-now-with-ai/)
Deployment is the bigger number
**1,000,000+**
robots deployed by Amazon across more than 300 facilities
Operator-reported · June 2025
**542,000**
industrial robots installed worldwide in 2024 alone
International Federation of Robotics · 2025 release
The startup opening is not “build every robot.” A small team can own one stubborn task, instrument its failures and make a measurable improvement that repeats across sites.
- [Verify Amazon’s fleet milestone →](https://www.aboutamazon.com/news/operations/amazon-million-robots-ai-foundation-model)
- [Read the IFR installation data →](https://ifr.org/ifr-press-releases/global-robot-demand-in-factories-doubles-over-10-years)
**How to read this:** the ≈700 and ≈200 figures are directional landscape estimates, while the martech figure counts solutions. Amazon’s total includes multiple robot types, and IFR counts industrial installations rather than humanoids. None of the figures alone is a market-size forecast or a claim about startup success.
The last centimetre · hands + actuation
## The robot often stops at the wrist.
Many robot bodies are sold or configured without a dexterous hand. The base, arm and end effector are often separate procurement decisions, so “buying the robot” may still leave the hardest contact problem—and a second SDK—to solve.
Separate is normal
### A hand is a subsystem, not an accessory.
LimX lists a fist-shaped end as standard on Luna and a five-finger hand as optional. Unitree publishes separate removal and installation guides for Dex3‑1 and Inspire hands on different G1 configurations. PSYONIC sells its Ability Hand with an open API for robot integration and shows it on Apptronik’s Apollo. These are concrete examples of a broader integration pattern: the body may provide a wrist interface while the hand, sensors, control electronics and policy data come from another supplier.
**Compatibility boundary:** a removable hand is not automatically a compatible hand. Confirm flange geometry, handedness, mass and inertia, power, communication, control mode and rate, collision model, safety limits, calibration, firmware and the training embodiment before ordering.
- [See Luna’s optional hand →](https://www.limxdynamics.com/en/products/luna/spec)
- [See G1 hand swap guides →](https://www.unitree.com/mobile/app/g1/)
- [Inspect PSYONIC’s robot API →](https://www.psyonic.io/robots)
> Visual: Hand integration layers
1. 01 **Wrist** Flange · cable route · mass · inertia
2. 02 **Hand** Fingers · transmission · force · speed
3. 03 **Touch** Taxels · force/torque · calibration · wear
4. 04 **Control** SDK · protocol · rate · safety limits
5. 05 **Policy** Retargeting · data · sim model · evaluation
- [Video: Wuji Hand 2](https://www.youtube.com/watch?v=7QYedp3ozjw)
Wuji Technology · Hand 2 Beta 1
### Twenty independently driven joints, exposed as a developer platform.
Wuji documents 20 active, back-drivable direct-drive rotary joints, a 1,000 Hz × 20-axis control rate, Ethernet, 12 V input and a 745 ± 10 g hand with soft body. The Python SDK, ROS 2 route and URDF, MJCF and USD assets make the integration surface unusually inspectable. The current documentation is also explicit that this is Beta 1 and that its external connector form will change; published durability language is not a substitute for your load, impact and lifecycle test.
**Active DoF** — 20
**Control** — 1 kHz × 20 axes
**Mass** — 745 ± 10 g
**Status** — Beta 1
- [Inspect product claims →](https://wuji.tech/en/hand2)
- [Read the current integration docs →](https://docs.wuji.tech/docs/en/wuji-hand/latest/overview/)
- [Video: ORCA Dexterity announces three new open source robotic hands](https://www.youtube.com/watch?v=WNtlUViSrPg)
ORCA Dexterity · open-source hand family
### Choose touch, simplicity or the full research platform.
ORCA’s announcement introduces three open-source routes: orcahand touch, orcahand lite and the classic orcahand. The classic hand is a 17-DoF tendon-driven platform with integrated tactile sensing, published design files, control code, bill of materials and assembly documentation. Open access makes the system easier to inspect, reproduce and repair; it does not make the three variants equivalent or remove the need to validate load, sensing, maintenance and licensing for your deployment.
**Family** — Touch · Lite · Classic
**Classic DoF** — 17
**Architecture** — Tendon-driven
**Artifacts** — CAD · code · BOM
- [Watch the ORCA announcement →](https://www.youtube.com/watch?v=WNtlUViSrPg)
- [Inspect the open platform →](https://www.orcahand.com/paper)
- [Video: Sharpa Wave dexterous robotic hand](https://www.youtube.com/watch?v=GcTUlOHvdOs)
Sharpa · Wave
### Make touch part of the control stack.
Wave is a one-to-one human-scale hand with 22 active degrees of freedom and a proprietary tactile array. Sharpa reports 0.02 N tactile sensitivity, more than 20 N fingertip force, greater than 4 Hz full-gesture speed and 2.5 million press-test cycles. It also publishes ROS 2, SDK, firmware, URDF, tactile-simulation and mechanical-CAD resources. Those figures are vendor tests without the complete protocols needed for cross-vendor comparison; validate payload, repeatability, tactile drift, wear, repair time and performance on your objects.
**Active DoF** — 22
**Touch** — 0.02 N claimed
**Fingertip** — >20 N claimed
**Assets** — ROS 2 · Isaac · CAD
- [Inspect specs + test claims →](https://www.sharpa.com/pages/wave)
- [Open software + CAD resources →](https://www.sharpa.com/pages/downloads)
- [Video: HW1 Launch & Feature Showcase | Gesture Platforms](https://www.youtube.com/watch?v=LPbB96fT_uw)
Gesture Platforms · HW1
### A lightweight hand aimed at desktop research and repair.
HW1 is an ESP32-S3 robotic-hand platform with 10 actively controlled degrees of freedom across 19 joints. Gesture states a mass below 500 g, repeatability within 1 mm, 100 Hz control, USB-C and Bluetooth Low Energy, plus motor-angle, current and temperature telemetry. The launch video also emphasizes replaceable fingers, accessible electronics and an offline desktop app. These are vendor specifications for a crowdfunding-stage, pre-delivery product—not independent lifetime or task evidence. Validate grasp payload, play, repeatability, heat, impact survival, spare-part supply, SDK maturity and delivery status before treating it as lab infrastructure.
**Active DoF** — 10 · 19 joints
**Controller** — ESP32-S3
**Mass** — <500 g claimed
**Status** — Pre-delivery
- [Inspect the current product page →](https://gestureplatforms.com/)
- [Read the independent technical summary →](https://www.cnx-software.com/2026/05/28/gesture-hw1-10-dof-esp32-s3-robotic-hand-with-high-dexterity-manipulation/)
- [Video: Introducing the mimic hand M1](https://www.youtube.com/watch?v=ikjPRgE8WLM)
mimic robotics · hand M1
### Match the data-capture hand to the robot hand.
mimic unveiled M1 and its U1 wearable in July 2026 as a full-stack dexterous-manipulation platform. The tendon-driven M1 has 15 active degrees of freedom across 21 joints and moves its primary actuators into the forearm. It is a current vendor launch—not yet independent durability or production evidence—so validate payload, backdrivability, collision recovery, retargeting, integration and availability on the intended task.
**Active DoF** — 15 · 21 joints
**Architecture** — Tendon-driven
**Data capture** — U1 wearable
**Status** — July 2026 launch
- [Inspect the launch + specifications →](https://www.mimicrobotics.com/blog/solving-dexterity-a-full-stack-approach)
- [Review the current company →](https://www.mimicrobotics.com/about)
- [Video: No Gears. No Belts. The Future of Robotics Starts Here](https://www.youtube.com/watch?v=WwxNc5Bh74E)
Genesis Advanced Technology · LiveDrive direct-drive actuator
### LiveDrive replaces the gearbox and belt in a Delta robot.
The video argues that mechanical motion has not advanced as quickly as robotics software, AI, sensing and vision. Genesis presents LiveDrive as a high-torque direct-drive actuator that removes the gearbox and belt, then demonstrates it powering a Delta robot for high-speed pick-and-place work. The company says the oil-free design reduces drivetrain complexity, maintenance and contamination risk while allowing denser layouts and precise control. Those are vendor claims: validate torque, heat, energy use, cycle time, repeatability, uptime and service life on the real workload.
Actuation lens · direct-drive Delta robot
- [Watch the full LiveDrive demonstration →](https://www.youtube.com/watch?v=WwxNc5Bh74E)
- [Jump to the direct-drive actuator →](https://www.youtube.com/watch?v=WwxNc5Bh74E&t=74s)
- [Jump to the Delta robot →](https://www.youtube.com/watch?v=WwxNc5Bh74E&t=127s)
- [Review the current company →](https://www.genesisadvancedtechnology.com/)
Mechanical fit
### Mount the real mass.
Check flange, cable bend, wrist workspace, self-collision, inertia, center of mass and payload after the hand and tool are attached.
Contact envelope
### Test the object distribution.
Measure fingertip and grasp force, speed, compliance, impact survival, wear, contamination and the smallest stable grasp on your objects.
Control surface
### Time the whole loop.
Verify voltage, current peaks, bus, command mode, feedback fields, update rate, latency, watchdogs and safe behavior after packet loss.
Learning surface
### Match simulation to hardware.
Pin SDK and firmware versions, inspect URDF/USD/MJCF assets, calibrate touch, retarget demonstrations and preserve failed grasps in evaluation.
**Evidence policy:** degrees of freedom count motion axes, not useful dexterity. Force, speed and tactile sensitivity use different fixtures and definitions across vendors. Ask for the test protocol, lifecycle curve, spare-finger or module process, repair turnaround and customer task data before comparing headline numbers.
The first deployment
## Start with a task you can count.
A narrow wedge gives the team a fixed environment, a repeatable dataset and a customer metric. Generality can be earned later by adding nearby tasks.
01 · Observe
### Shadow the work.
Record task frequency, handling variation, travel, exceptions, safety controls and the human skills that make recovery look easy.
**Output: task + exception map**
02 · Price
### Find the costly constraint.
Quantify labour hours, injury exposure, downtime, scrap, throughput, night coverage and the cost of doing nothing.
**Output: value per successful cycle**
03 · Bound
### Shape the environment.
Standardize bins, lighting, floor markers, approach angles, fixtures or handoff points before demanding more intelligence.
**Output: operating design domain**
04 · Prototype
### Teleoperate before autonomy.
Use remote operation or a scripted controller to prove reach, payload, cycle time and customer workflow before training a policy.
**Output: working data collector**
05 · Learn
### Simulate, collect and evaluate.
Version scenes, demonstrations, checkpoints and tests. Include recoveries and off-nominal states—not only polished success episodes; reuse the [model evaluation loop](https://isaiuseful.com/training-models.html.md#loop) for every learned component.
**Output: reproducible policy evidence**
06 · Pilot
### Keep a human recovery path.
Run inside a fenced scope with stop controls, incident logging and an operator who can recover the task without improvisation.
**Output: field reliability + economics**
Good first wedge
### One object family, one site, one shift.
Examples: machine tending for a named part, tote movement on a mapped route, visual inspection at a fixed station, or cleaning one repeatable floor type.
Bad first wedge
### “Replace any worker anywhere.”
No stable task distribution, no realistic acceptance set, no bounded safety case and no credible denominator for unit economics.
System architecture
## The robot is a chain of assumptions.
Choose the simplest component at every layer that preserves the field requirement. A spectacular policy cannot repair the wrong gripper, missing stop circuit or brittle site integration.
Job + environment
### Operating design domain
Task, objects, people, floor, lighting, weather, network, shift and allowed exceptions.
Acceptance starts here.
Body + tooling
### Embodiment
Wheels, legs, arm, gripper, payload, reach, speed, durability and maintainability.
Use a humanoid only when its form earns access.
Sensing + compute
### Perception at the edge
RGB, depth, LiDAR, force, proprioception, synchronization, bandwidth and thermal budget.
Redundancy is a safety and reliability decision.
Planning + control
### Deterministic and learned layers
State estimation, navigation, motion planning, low-level control, learned policy and safety supervisor.
Do not let a VLA bypass hard limits.
Data + simulation
### Learning loop
Teleoperation, scene generation, demonstrations, failure capture, policy training and regression evaluation.
Version the environment with the model.
Fleet + service
### Production system
Deployment, observability, remote assist, spares, updates, access control, incident response and customer support.
The service margin lives here.
- [Robot Operating System →](https://www.ros.org/)
- [Gazebo simulation →](https://gazebosim.org/)
- [Drake planning and control →](https://drake.mit.edu/)
- [LeRobot real-world loop →](https://huggingface.co/docs/lerobot/main/en/getting_started_real_world_robot)
- [NVIDIA Isaac Sim documentation →](https://docs.nvidia.com/isaacsim/latest/index.html)
Model + edge computer
## Robostral and Thor solve different layers.
One is a task-specific navigation model; the other is a compute platform. A product still needs sensors, localization, control, stop logic, integration and a measured recovery loop. Use the [local-model memory guide](https://isaiuseful.com/local-models.html.md#memory) to shortlist adjacent edge candidates before measuring them on-device.
- [Video: Introducing Robostral Navigate](https://www.youtube.com/watch?v=7dpLB9NoY1A)
Embodied navigation · Mistral
### Language instruction → a route through the world
Mistral presents an 8B navigation model trained in simulation that uses one RGB camera. Its reported R2R‑CE success is benchmark evidence for the named task—not proof that any robot can safely navigate any site.
- [Read Mistral’s technical account →](https://mistral.ai/news/robostral-navigate/)
- [Video: Getting Started with the NVIDIA Jetson AGX Thor Developer Kit for Physical AI](https://www.youtube.com/watch?v=iYT2haVIgSM)
Edge compute · NVIDIA Developer
### Put a larger multimodal stack at the edge
Jetson AGX Thor combines 128 GB of memory, a Blackwell GPU and extensive robot I/O in a 130 W-class developer platform. Peak FP4 compute is not application latency; test the exact model, sensors, thermal envelope and control loop.
- [Open the current developer guide →](https://docs.nvidia.com/jetson/agx-thor-devkit/user-guide/latest/)
- [Verify NVIDIA’s published specifications →](https://nvidianews.nvidia.com/news/nvidia-blackwell-powered-jetson-thor-now-available-accelerating-the-age-of-general-robotics)
Navigation policy
### Robostral Navigate
**Input** — RGB history + language instruction
**Output** — Image point/orientation or local displacement
**Reported scale** — 8B · 2.4M simulated trajectories · 350k scenes
**Evidence** — Vendor report on R2R‑CE and office demonstrations
Reproduce the named benchmark and then test your camera placement, route geometry, people, glare, obstacles and recovery states. A model score is not a site safety case.
- [Inspect the release evidence →](https://mistral.ai/news/robostral-navigate/)
Robot computer
### Jetson AGX Thor
**Memory** — 128 GB unified LPDDR5X
**Vendor peak** — Up to 2,070 FP4 TFLOPS
**Developer power mode** — Up to 130 W; supplied adapter constraints apply
**I/O** — USB, camera, Ethernet and QSFP28 routes
Start from the current JetPack release, pin containers and measure end-to-end sensor-to-action latency under sustained heat. The developer kit is not the final production carrier or certification plan.
- [Open the current Thor guide →](https://docs.nvidia.com/jetson/agx-thor-devkit/user-guide/latest/)
Learned layer
**Interpret, perceive, propose**
Language, scene grounding, task planning and candidate actions.
Supervised control
**Validate, constrain, execute**
Workspace limits, collision checks, speed limits, watchdogs and stop circuits.
Field evidence
**Log, review, improve**
Interventions, near misses, task failures, recovery time and accepted cycles.
Practitioner experience · sentdex
## Follow one humanoid from unboxing to learned locomotion.
This series is valuable because it exposes the integration work between the product film and a custom behavior: networking, LiDAR, SLAM, arm and hand control, external compute, simulation and reinforcement learning.
- [Video: Unboxing the Unitree G1 EDU Humanoid](https://www.youtube.com/watch?v=pPTo62O__CU&t=2859s)
01 · Hardware baseline
**Unboxing the Unitree G1 EDU Humanoid**
*Inventory what arrives, what is EDU-specific and what still needs integration.*
- [Video: LiDAR, SLAM, navigation and control](https://www.youtube.com/watch?v=sJYlJlIEBpg&t=2251s)
02 · Mobility stack
**LiDAR, SLAM, navigation and control**
*Connect perception, mapping and the first controllable movement loop.*
- [Video: Moving the arms and hands](https://www.youtube.com/watch?v=Uc1nhT8beTU)
03 · Manipulation
**Moving the arms and hands**
*See where command interfaces meet joint limits and end-effector reality.*
- [Video: A bigger brain for the G1](https://www.youtube.com/watch?v=cmnJhOWp2z4)
04 · External compute
**A bigger brain for the G1**
*Trace the networking and architecture questions behind off-board intelligence.*
- [Video: Vibe coding the Inspire robot hands](https://www.youtube.com/watch?v=MeHWIXLV3Zo)
05 · Hand interface
**Vibe coding the Inspire robot hands**
*Turn a vendor interface into a small, testable manipulation experiment.*
- [Video: Make a robotic hand crawl](https://www.youtube.com/watch?v=57cPmzwCqd4)
06 · Creative control
**Make a robotic hand crawl**
*A compact study in actuation, iteration and unexpected embodiments.*
- [Video: Reinforcement learning with G1](https://www.youtube.com/watch?v=wiIUF9pIDYw&t=93s)
07 · Policy training
**Reinforcement learning with G1**
*Move from built-in behavior to a reproducible simulation and training route.*
- [Video: Train a G1 to walk](https://www.youtube.com/watch?v=FGnAeUXRZ4E)
08 · Sim to real
**Train a G1 to walk**
*Connect a learned locomotion policy to the physical deployment boundary.*
Vendor baseline
### Know the exact G1.
Unitree lists configurations ranging from 23 to 43 joint motors. EDU options, hands, compute and sensors change the software surface and the price.
- [Check current G1 configurations →](https://www.unitree.com/g1/)
Real robot interface
### Pin the SDK and firmware.
The official SDK2 route uses DDS interfaces for G1 and other Unitree robots. Confirm the exact messages and services available on your firmware before building control around them.
- [Inspect Unitree SDK2 →](https://github.com/unitreerobotics/unitree_sdk2)
Learning loop
### Rehearse before deployment.
Unitree publishes MuJoCo, Isaac Lab and reinforcement-learning routes plus a LeRobot bridge for G1 data, training and real-world evaluation.
- [Unitree MuJoCo →](https://github.com/unitreerobotics/unitree_mujoco)
- [Unitree RL Lab →](https://github.com/unitreerobotics/unitree_rl_lab)
- [Unitree LeRobot →](https://github.com/unitreerobotics/unitree_lerobot)
Deployment gate
## Measure the whole service, not the demo.
Choose thresholds before the pilot. Report a distribution and its failure denominator—not only the cleanest successful video. The site’s [evidence boundary](https://isaiuseful.com/evidence.html.md#claim-boundary) explains why a strong result on one task must stay scoped to that task.
Task
### Successful cycles
**Successes ÷ all attempted cycles**
Segment by object, route, lighting, operator and exception type.
Intervention
### Human recovery
**Interventions per operating hour**
Include remote assist, resets, falls, stuck states and manual completion.
Time
### Useful throughput
**P50 / P95 cycle + recovery time**
Compare with the actual human or machine process—not a lab ideal.
Reliability
### Availability
**Uptime, MTBF and repair time**
Track batteries, sensors, joints, networking, software and consumables separately.
Safety
### Leading indicators
**Stops, near misses and limit violations**
Treat “no injury” as insufficient when the pilot is small.
Economics
### Cost per accepted cycle
**Robot + service + people + site changes**
Include financing, support, spares, teleoperation and customer integration.
Safety boundary
### Risk assessment belongs to the robot system.
The end effector, program, power, sensors, communication interfaces, cell, people and maintenance workflow all matter. OSHA’s industrial-robot guidance distinguishes manufacturer, integrator and user responsibilities and points to task-based risk assessment and safeguarding.
- [Read OSHA’s robot-system guidance →](https://www.osha.gov/otm/section-4-safety-hazards/chapter-4)
European route
### Plan conformity before the hardware freezes.
The EU Machinery Regulation replaces the Machinery Directive and applies from January 2027. An AI system used as a safety component of machinery can also enter the AI Act’s high-risk route when the legal criteria are met. Scope the actual product with qualified safety and legal specialists.
- [EU Machinery Regulation →](https://eur-lex.europa.eu/legal-content/en/ALL/?uri=CELEX:32023R1230)
- [AI Act Article 6 classification →](https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-6)
Company layer · Europe
## Build the cap table for a hardware journey.
Robotics crosses borders early: suppliers, certification, pilots, investors and hires rarely sit in one Member State. Company structure is infrastructure—but it does not replace local tax, labour, product-safety or operating obligations.
EU—INC
### One European standard is a proposal, not today’s incorporation route.
EU‑INC is a grassroots campaign for a pan-European corporate standard. Its July 2026 position paper argues for free choice of registered office, one central registry, broad company access, standardized stock options, and local labour law and taxes tied to real activity.
This is an advocacy position. Verify the final law, national implementation and your actual incorporation, employment, tax, IP and fundraising facts with qualified advisers.
Before the first institutional round
- IP assignment from every founder, contractor and university partner
- Hardware, dataset, model and open-source licence inventory
- Supplier terms, export controls and product-liability allocation
- Employee option plan that works where the team actually lives
- Pilot contracts that separate experiments from production commitments
- [Read the EU‑INC position paper](https://www.eu-inc.org/position-paper)
Prototype Capital starter pack
## Use the link list as a route, not a reading pile.
Every substantive external resource linked in Prototype Capital’s “Why you should start a robotics startup” post is organized below. The post is an investor’s editorial case for the opportunity; the primary courses, tools and documentation are the stronger sources for implementation details.
- [Video: Why you should start a robotics startup](https://www.youtube.com/watch?v=FqfTQFuSalY)
Opportunity thesis · Prototype Capital
### Why 2026 may be a strong moment to start
The video argues that cheaper components, improving vision-language-action models and macro demand are expanding the design space. Treat the market counts and timing as an investor thesis; validate the named customer task and procurement reality yourself.
- [Read the source post →](https://updates.prototypecap.com/p/why-you-should-start-a-robotics-startup)
- [Prototype Capital’s YouTube channel →](https://www.youtube.com/@prototypecap?sub_confirmation=1)
Jump within the embedded video
16 chapters
- [Video chapter: 00:00 Intro](https://www.youtube.com/watch?v=FqfTQFuSalY)
- [Video chapter: 00:34 Robotics are exploding](https://www.youtube.com/watch?v=FqfTQFuSalY&t=34s)
- [Video chapter: 01:58 Why start in 2026?](https://www.youtube.com/watch?v=FqfTQFuSalY&t=118s)
- [Video chapter: 03:56 Component costs](https://www.youtube.com/watch?v=FqfTQFuSalY&t=236s)
- [Video chapter: 04:42 The “spaghetti moment”](https://www.youtube.com/watch?v=FqfTQFuSalY&t=282s)
- [Video chapter: 06:11 The unsettled stack](https://www.youtube.com/watch?v=FqfTQFuSalY&t=371s)
- [Video chapter: 06:53 The data problem](https://www.youtube.com/watch?v=FqfTQFuSalY&t=413s)
- [Video chapter: 07:46 Macro tailwinds](https://www.youtube.com/watch?v=FqfTQFuSalY&t=466s)
- [Video chapter: 08:40 Robotics as the new SaaS](https://www.youtube.com/watch?v=FqfTQFuSalY&t=520s)
- [Video chapter: 09:14 Structured environments](https://www.youtube.com/watch?v=FqfTQFuSalY&t=554s)
- [Video chapter: 10:05 Global niches](https://www.youtube.com/watch?v=FqfTQFuSalY&t=605s)
- [Video chapter: 10:32 General vs vertical](https://www.youtube.com/watch?v=FqfTQFuSalY&t=632s)
- [Video chapter: 11:54 Small-team speed](https://www.youtube.com/watch?v=FqfTQFuSalY&t=714s)
- [Video chapter: 12:32 What about humanoids?](https://www.youtube.com/watch?v=FqfTQFuSalY&t=752s)
- [Video chapter: 14:52 Open opportunities](https://www.youtube.com/watch?v=FqfTQFuSalY&t=892s)
- [Video chapter: 15:37 Summary](https://www.youtube.com/watch?v=FqfTQFuSalY&t=937s)
Start here
### Touch the learning loop.
- [LeRobot robot-learning tutorial →](https://huggingface.co/spaces/lerobot/robot-learning-tutorial)
- [Mythbusting physical AI →](https://planet-a.medium.com/robots-in-the-real-world-mythbusting-physical-ai-f9eb1688ee35)
- [The physical-AI deployment gap →](https://www.a16z.news/p/the-physical-ai-deployment-gap)
- [The new labour economy thesis →](https://newsletter.semianalysis.com/p/america-is-missing-the-new-labor-economy-robotics-part-1)
Learn deeply
### Build the foundations.
- [Springer Handbook of Robotics →](https://link.springer.com/referencework/10.1007/978-3-319-32552-1)
- [MIT Robotic Manipulation →](https://manipulation.csail.mit.edu/)
- [MIT Underactuated Robotics →](https://underactuated.csail.mit.edu/)
- [NVIDIA Robotics learning path →](https://www.nvidia.com/en-us/learn/learning-path/robotics/)
Build something
### Pick one implementation spine.
- [Robot Operating System →](https://www.ros.org/)
- [Gazebo →](https://gazebosim.org/)
- [Drake →](https://drake.mit.edu/)
- [Awesome Robotics index →](https://github.com/kiloreux/awesome-robotics)
Keep current
### Follow deals and practice.
- [Robotics Roundup →](https://roboticsroundup.substack.com/)
- [Robotics for Software Engineers →](https://newsletter.pragmaticengineer.com/p/robotics-for-software-engineers)
- [Lukas M. Ziegler →](https://www.linkedin.com/in/zieglerr/)
- [Lukas Ziegler’s newsletter →](https://ziegler.substack.com/)
Track the ecosystem
### Add funding, standards and workforce views.
- [Paulina Szyzdek →](https://www.linkedin.com/in/paulina-szyzdek)
- [Paulina Szyzdek’s newsletter →](https://paulinaszyzdek.substack.com/)
- [Aaron Prather →](https://www.linkedin.com/in/amprather)
The practical answer
Choose one physical job. Instrument every failure. Make reliability earn scale.
- [Define the wedge](#route)
- [Assemble the stack](#stack)
- [Set the deployment gate](#scoreboard)
- [Browse AI Now engineering talks](https://isaiuseful.com/training-models.html.md#ai-now-summit)