FetchMan Tests Sim-to-Real Humanoid Picking

Single-source brief: FetchMan reports a simulation-trained policy on a real Unitree G1, but independent confirmation was not supplied.

What Changed

FetchMan targets visual tasks that require a humanoid to walk, balance, and pick objects.

The authors say common data-heavy methods need many demonstrations. That need grows when robots must move their full bodies.

Their pipeline trains on more than 150,000 simulated scenes. It first clones synthetic demonstrations. It then uses reinforcement learning to improve the policy.

The paper says synthetic behavior cloning hit a performance ceiling. More training data did not remove that limit.

The team then refined the cloned policy with Flow-GRPO. It used one sparse reward signal.

Reported Robot Result

According to the arXiv paper, the team deployed a single-object reach-and-pick policy zero-shot on a Unitree G1.

“Zero-shot” means the authors report no added real-world training for that deployment.

The robot walked to and grasped a target in unseen scenes. The paper reports a 73.3% success rate.

The authors also released FetchMan-Bench, a simulation benchmark. They describe multi-object training as an early step toward broader humanoid policies.

Deployment Reality

This is a research-stage result, not a confirmed commercial deployment.

The reported task is narrow: reach and pick one target. The supplied evidence does not state runtime, payload, safety limits, or recovery behavior.

It also does not provide independent validation. The real-world result comes from the paper’s authors and arXiv record.