SafeHumanoid: VLM-RAG-Driven Impedance Control of Humanoid Robot

Safe Human–Robot Interaction in shared spaces requires robots to adapt compliance and speed to scene context and human proximity. Standards for collaborative operation specify limits on speed, force/pressure, and stiffness during contact or near-contact operation, motivating dynamic impedance rather than fixed gains. We present an egocentric vision pipeline for the Unitree G1 humanoid that retrieves task- and context-appropriate parameters via a Vision-Language Model (VLM) with Retrieval-Augmented Generation (RAG), and integrates inverse kinematics (IK) to produce joint references together with feed-forward torques for gravity compensation. Concretely, the system maps live head-camera frames to (Kp, Kd, v); IK provides (qref, qref, τff); and the low-level controller apply it at each joint. This enables joint-impedance scheduling and speed regulation in low-dynamic settings. We evaluate the system on dual-manipulation tabletop tasks with and without human presence in the scene. Compared to fixed-gain baselines, the VLM-RAG policy yields smooth speed reductions near humans, and maintained task success under environmental change. While the off-board perception–retrieval loop introduces noticeable delay that restricts use in highly dynamic interaction, the prototype demonstrates that semantic grounding of impedance remains a viable direction for safer humanoid collaboration.