New RAPT model enhances humanoid robot safety in sim-to-real transfers

Researchers have developed RAPT (Recurrent Anomaly Probabilistic Trajectory Model), a new system designed for detecting out-of-distribution states and diagnosing failures in humanoid robots during sim-to-real deployment. This lightweight, self-supervised model operates at 50 Hz and learns nominal robot behavior from simulation data to predict deviations in real-world execution. RAPT aims to provide calibrated, per-dimension predictive-deviation signals, enabling detection under strict false-positive constraints and localizing when and where execution diverges from nominal behavior. For post-hoc diagnosis, RAPT integrates temporal saliency, joint-kinematic summaries, and LLM-based semantic reasoning to classify failure causes in a zero-shot manner. AI

IMPACT Enhances safety and reliability for humanoid robots transitioning from simulation to real-world deployment.

RANK_REASON Research paper detailing a new model for robot control safety. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• DagsHub
• Hugging Face
• Humphrey Munning
• Isaac Lab
• Recurrent Anomaly Probabilistic Trajectory Model

• paper
• safety
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →