Bimanual Robot Manipulation via Multi-Agent In-Context Learning

The article presents BiCICLe (Bimanual Coordinated In-Context Learning), a novel framework that enables standard language models to execute few-shot bimanual manipulation tasks without fine-tuning. By framing the control problem as a multi-agent leader-follower scenario, BiCICLe effectively decouples the action space into sequential single-arm predictions, achieving a 70.5% average success rate across 13 tasks from the TWIN benchmark, outperforming existing training-free and many supervised methods. This advancement is significant for practitioners as it allows for improved robot manipulation capabilities using off-the-shelf LLMs, facilitating broader applications in robotics without the need for extensive hardware-specific retraining.