🤖 AI Summary
Existing robotic simulation modeling relies heavily on hand-crafted kinematic/dynamic models and large-scale real-world data, limiting autonomy and generalizability. Method: We propose an end-to-end framework for autonomous construction of differentiable simulators, enabling robots to jointly identify their morphology, kinematics, and dynamics solely through active interaction and multimodal perception (vision and proprioception). Our approach integrates deep reinforcement learning, neural radiance fields (NeRF) for geometric representation, differentiable physics engines, and online system identification—without requiring prior dynamical knowledge or manual modeling. Contribution/Results: To our knowledge, this is the first method enabling closed-loop self-identification and policy pre-execution in simulation. Evaluated on a physical robotic arm, it achieves significantly higher dynamic fidelity: 47% reduction in action prediction error and a 3.2× improvement in task planning success rate—breaking the paradigm of manual modeling and extensive real-data dependency.
