A humanoid robot and a human partner learn to hand over objects — entirely by self-play reinforcement learning
A Unitree G1 humanoid takes a parcel from a simulated human partner's hand in a household apartment. Nothing in these clips is scripted. The robot and the partner were both trained from scratch, against each other, with self-play PPO: no demonstrations, no pretraining, no hand-designed handover motion.
Fine-Tuning OpenVLA-OFT for Household Manipulation Task
We fine-tuned OpenVLA-OFT on 100 expert demonstrations collected in an Isaac Sim kitchen environment, enabling a Franka robot to pick up a cup, transport it to the sink target region, and release it. The dataset contains 34,237 timesteps, and every demonstration was independently verified through action replay in simulation. The policy takes two RGB observations—from a front-side camera and a wrist-mounted camera—together with an 8-dimensional proprioceptive state and a language instruction. It predicts a 7-dimensional action consisting of end-effector translation deltas, rotation deltas, and a gripper command. In preliminary closed-loop evaluations, the fine-tuned policy performed well and successfully executed the complete grasping, transporting, and releasing sequence.
- Zhiyuan Gao
Email: gao@uni-bremen.de
Profile: Zhiyuan Gao
- Mohammad Khoshnazar
Email: khoshnazar@uni-bremen.de
Profile: Mohammad Khoshnazar
- Yanxiang Zhan
Email: yanxiang@uni-bremen.de
Profile: Yanxiang Zhan