RoboMirror: Bridging Human Intent and Robot Capability Through Self-Reflective Motion Adaptation

Abstract

Current humanoid robot control paradigms place the burden of feasibility assessment on human operators, who must carefully design commands within perceived robot limitations.  This constraint significantly hinders practical deployment and limits the expressiveness of robot behaviors.  This study proposed an inverting paradigm: rather than constraining operator inputs, robots should autonomously evaluate their capacity to execute commanded motions and intelligently adapt references to align with their physical constraints and learned skills. This study introduced the Performance Prediction Network (PPN), a transformer-based architecture that forecasts execution quality for arbitrary reference trajectories by analyzing both the commanded motion sequence and current robot state.  Given a high-level task specification, our framework synthesizes multiple viable motion candidates and employs PPN to rank them across six dimensions:  collision avoidance, kinematic feasibility, dynamic stability, trajectory smoothness, and goal satisfaction. This ranking enables autonomous selection of the most suitable reference motion before execution begins.  Our complete system integrates motion generation, kinematic retargeting, and learned control policies with PPN-guided adaptation, creating a closed-loop framework where robots reason about their own limitations.  Validated on 100,000 diverse human motions span walking, running, jumping, and acrobatic maneuvers, PPN achieves 99.14% accuracy in predicting imminent failures while maintaining low prediction error across all performance metrics. In deployment, our system successfully prevents 62% of anticipated falls by autonomously modifying commanded references, demonstrating that explicit capability modeling enables safer and more reliable humanoid control without sacrificing behavioral diversity.