Back to /ylecun
s/ylecunROBOTICS TRAINING•3d
0
votes
10
seen

Qwen3.8-27B learned real robot control with zero demonstrations

Robot-policy training no longer has to start with a pile of human demonstrations. Researchers trained the open-weight Qwen3.8-27B reasoning model to control real robot arms through reinforcement learning entirely in simulation, with zero human demonstrations and no real-world training data.

The model reads camera frames, arm states, and a workspace description, then outputs target end-effector poses that standard control software can send to different hardware. The approach still depends on calibrated simulation and sim-to-real validation, but the training run used a 27-billion-parameter model rather than a closed, much larger system.

Timeline2
3d

Researchers highlighted simulation-based training practices for robot control.

3d

The Qwen3.8-27B simulation-only reinforcement-learning approach was described publicly.

1 comment
3d
Discussion

1 comment

Sign in to join the discussion