RoboMAN: Human-like Compliance Manipulation for Electronics Assembly via an Online Memory-Augmented Network
Workshop Statement: Our work on RoboMAN directly addresses several challenges outlined in the workshop theme, particularly the use of large-scale data and models for human-centered robotics. By collecting a rich, high-frequency force-motion dataset through a bilateral teleoperation system, we leverage both human expertise and machine learning to build robust, data-driven compliance policies. This approach aligns closely with the workshop’s emphasis on identifying and acquiring high-quality robot interaction data, as well as ensuring that the learned models faithfully capture the subtleties of human behavior. Furthermore, our online memory-augmented framework facilitates adaptive fine-tuning of model parameters, a capability that mirrors the broader push in embodied AI toward continuous learning from diverse and evolving real-world inputs.
In addition, our focus on safe and precise force regulation in electronics assembly contributes to the workshop’s goals of AI safety and alignment in physically interactive settings. By grounding our framework in multimodal force and motion feedback, we minimize risks associated with robot-human and robot-environment interactions while maintaining high assembly success rates. This work therefore showcases how large-scale, human-centered data collection and adaptive model training can help advance state-of-the-art embodied AI systems that are both safety-aware and capable of intricate, contact-rich tasks.
Keywords: Human-like compliance control, Online memory-augmented network, Bilateral teleoperation system
TL;DR: This paper proposes RoboMAN, an online memory-augmented compliance framework that leverages a multimodal force-motion dataset from bilateral teleoperation to achieve robust, human-like compliance in precision electronics assembly tasks.
Abstract: Traditional compliance control methods face limitations in precision assembly due to delicate contact transitions and rigid parameter tuning. This paper introduces an online memory-augmented compliance learning framework RoboMAN to achieve human-like compliance control in precision electronics assembly tasks. This framework is trained on a 6-DOF force-motion dataset collected via our developed bilateral teleoperation system. Experimental evaluations on four representative electronics assembly tasks demonstrate RoboMAN’s superiority in memory efficiency (48% GPU utilization), training speed (65.25s per epoch), inference latency (0.25s per batch), task success rates (up to 98%), while demonstrating robust dynamic force adaptability during task execution. This work establishes a new paradigm for adaptive robotic compliance control by bridging biological compliance principles with efficient robotic execution, offering a scalable solution in precision contact-rich environments.
Submission Number: 17
Loading