Berkeley's MonoDuo uses one robot arm to train bimanual policies with 70% success

MonoDuo addresses a fundamental bottleneck in robotics: bimanual robots are rare and expensive, limiting the data needed for policies that coordinate two arms. The UC Berkeley team (Bajamahal, Chen, Lin, Ma, Malik, Goldberg) proposes a clever workaround: use a single-arm robot — common in labs — teleoperated by a human who simultaneously performs the other side of the task. By swapping roles, they collect paired data covering both sides. RGB-D observations from wrist and fixed cameras are augmented with hand pose estimation, segmentation, and inpainting to create synthetic demonstrations that ground in real robot kinematics. This allows training bimanual policies without ever needing a dual-arm robot during data collection.

Evaluated on five challenging tasks — box lifting, backpack packing, cloth folding, jacket zipping, and plate handover — MonoDuo achieves up to 70% success in zero-shot deployment on unseen bimanual configurations. When few-shot finetuning is applied with only 25 target robot demonstrations, success rates jump 65-70% compared to training from scratch. This demonstrates efficient knowledge transfer from single-arm data to bimanual policies, drastically reducing the need for expensive bimanual hardware. The work, accepted at ICRA 2026, could democratize bimanual manipulation research by enabling labs with only single-arm robots to contribute meaningfully.