Finding
Robots now learn tasks from human demonstrations and simulation instead of hand-written programs, so real-world data has become the bottleneck
Industrial arms were traditionally programmed point by point. The new approach trains a neural network on demonstrations - people teleoperating the robot - then refines it in simulation with reinforcement learning, and keeps improving it from data collected in deployment. Vision-language-action models such as Nvidia's GR00T, Figure's Helix and Physical Intelligence's models let one network take camera input and an instruction and output motor commands.
Data is the scarce input. Driving models learn from millions of cars; robots have far smaller fleets, so companies collect teleoperation hours, use simulation heavily and share data across robot types. Unitree, AgiBot and other Chinese makers have open-sourced models to build ecosystems. The robots that improve fastest will be the ones deployed in volume, which is the same loop that drives autonomous vehicles.
Methods
Robot learning methods
| Method | How it works | Strength |
|---|---|---|
| Imitation learning | Copy human teleoperated demonstrations | Fast to teach new tasks |
| Reinforcement learning | Trial and error with rewards, mostly in simulation | Finds robust behaviors |
| Simulation (sim-to-real) | Train in physics simulators, transfer to hardware | Cheap, safe practice |
| Foundation models | One model across tasks and robots | Generalization |
| Fleet learning | Collect data from deployed robots | Continuous improvement |
Sources
Sources and scope
- Scope
- AI in Robotics
- Sources
- Nvidia; Figure; Physical Intelligence; Unitree
- Updated
- October 2026