New AI method auto-improves physics for robot simulation objects

A new thesis from Anh-Quan Pham tackles a foundational bottleneck in robot learning: the gap between geometrically rich 3D object datasets and the physical properties needed for stable simulation. The author introduces the concept of "interaction-readiness," a quantitative framework that decomposes object quality into measurable components—revealing failure modes like unstable joints or unrealistic friction that standard evaluations miss. The core contribution is a multi-modal, simulator-in-the-loop approach that combines geometric, visual, and semantic cues to infer physical parameters such as mass, inertia, and joint limits from incomplete 3D assets. It then iteratively adjusts these parameters based on simulator feedback until the object behaves reliably under manipulation.

Experiments span a diverse set of articulated objects (cabinets, scissors, doors) and manipulation tasks (pushing, opening, grasping). Results show that interaction-readiness directly correlates with simulation stability and downstream policy performance: policies trained on refined objects achieve higher success rates and more consistent behavior than those on unrefined assets. This work provides a practical pipeline for scaling simulation-ready object creation without manual tuning, making it highly relevant for robotics labs and sim-to-real transfer pipelines.