Teaching robots now takes data operations, simulation and careful testing. These figures and dated updates show where training data comes from, how fast robots learn new tasks and how results are checked.
What moved in robot training this year.
270,000 hours of real-world manipulation data behind Generalist's GEN-0 model (Generalist AI, 2025)
$100M+ a year that robotics companies spend on real-world movement data (PYMNTS, 2026)
<200 examples to adapt Gemini Robotics On-Device 2 to a new two-arm robot (The Robot Report, 2026)
Runway unveils Praxis-1, a robot control model built on video pretraining
· Research
Video AI company Runway introduced Praxis-1, an open-weight world action model that turns large-scale video pretraining into robot control instead of relying mainly on scarce robot demonstrations. Runway says testing policies inside its world model predicted real-world results with 0.95 correlation. Early partners are testing it, and a public release is planned.
Why it matters: If simulated tests reliably predict real results, teams can screen new robot skills before spending time on hardware. Ask any vendor how its virtual evaluations compare with measured results on site.
Innodata opens a motion-capture lab to supply and verify humanoid training data
· Launch
Data company Innodata opened a motion-capture lab in Ridgefield Park, New Jersey, using Vicon infrared cameras that track movement with sub-millimeter accuracy. It sells ready-made and custom motion datasets for humanoids and other robots, retargets data across platforms, and offers independent performance evaluation and safety benchmarking.
Why it matters: Third-party testing gives buyers a check on maker claims. Before a pilot, ask whether a robot's skills were measured independently and whether you can rerun the same tests at your site.
Japanese consortium opens a school where VR operators teach humanoids factory work
· Deployment
J-HRTI, a consortium of five companies including Tsumura and Yamazen, opened a facility in Narashino, Chiba Prefecture, that can hold up to 40 Chinese-made humanoids. Operators wearing VR headsets control the robots in mock work areas to collect training data, aiming to deploy them on simple tasks such as moving materials by March 2027.
Why it matters: Pooling teleoperation data spreads the cost of teaching robots across several firms. If you join or buy into such a program, agree who owns the recorded data and the skills trained on it.
Generalist's GEN-1.5 learns simple tasks from seconds of human demonstration
· Research
Generalist AI said its GEN-1.5 model reached 59% average success on ten short manipulation tasks after a single demonstration of 3 to 12 seconds, with no retraining. With about 50 demonstrations, or five minutes of data per task, and brief fine-tuning, average success rose to 83%.
Why it matters: Fewer demonstrations per task would cut the cost of teaching a robot site-specific work. Treat lab success rates as a starting point and measure them again on your own objects and layout.
Reimagine Robotics debuts robots that factory staff teach by showing and correcting
· Launch
London- and Sydney-based Reimagine Robotics, founded by a former Google DeepMind leader, emerged from stealth with pre-seed funding. Workers show its robots a task, watch them try and correct mistakes in real time, with no programmer. Early projects include tending 3D printers overnight at a plastics plant and a three-robot hard-drive disassembly cell.
Why it matters: When frontline staff can teach and correct a robot, changing a task no longer waits for an engineer. THG Academy courses prepare operators for this kind of hands-on robot coaching.
Leju's Jinan training center puts 45 humanoids to work generating skill data
· Milestone
Robot maker Leju Robot runs a 4,000-square-meter humanoid training center in Pingyin county, Jinan, where 45 humanoids practice in 11 real-world scenarios, from industrial sorting to kitchens and hotel tasks. Operators use VR gear and sensor controllers to transfer their movements to the robots, producing more than 3 million data entries a year.
Why it matters: Training centers that stage real work settings now act as factories for robot skills. Ask makers where a skill was learned and how closely that setting matches your own site.
People also search for humanoid robot training news, robot training data, teleoperation data collection, robot data factory, humanoid robot training center and robot learning from demonstration.