FetchMan teaches humanoid robots to walk and grasp via 150K simulated scenes

Humanoid robots that can both walk and manipulate objects in unfamiliar environments have long been a robotics goal, but collecting enough real-world demonstrations is impractical. In the paper FetchMan, Omar Rayyan and colleagues tackle this by learning entirely from simulated experiences, then transferring the policy to the real world. Their end-to-end pipeline uses 150,000 scenes to train the robot, avoiding the data-hungry bottleneck of physical trials.

The training recipe is notable for its three-stage design. First, synthetic behavior cloning provides a coarse initial policy. But cloning alone hits a low performance ceiling regardless of data volume, so the team refines it with Flow-GRPO reinforcement learning using a single sparse reward. This RL refinement is the key to breaking through and achieving real-world competence, which pure cloning cannot match. The resulting policy, FetchMan, is first evaluated in the FetchMan-Bench simulation benchmark they release, then deployed zero-shot on a real Unitree G1 humanoid.

Results are promising: the single-object reach-and-pick policy achieves a 73.3% success rate across unseen scenes in the real world. The method also extends to multi-object training, marking an early step toward more generalist loco-manipulation policies at this data scale. For robotics researchers, this demonstrates that simulation-driven RL can transfer robustly to humanoid platforms, potentially accelerating practical deployment in warehouses, homes, and other unstructured settings.