Why robots still fall through the chair — and how a new framework teaches them not to

A person sits in a chair. The 3D model looks fine. In a physics simulator, the chair gives way and the body clips straight through.

That gap between "looks right" and "behaves right" is what HSImul3R from Daxiao Robot, NTU S-Lab and Shanghai AI Lab tries to close. The paper was accepted to ECCV 2026, a top computer-vision conference, and was covered this week by 量子位 (QbitAI).

The framework puts a simulator inside the training loop, not just at the end. It runs both ways: forward, it nudges a person's motion to respect contact and gravity; backward, it adjusts the 3D scene until physics says the pose holds.

The team built a benchmark, HSIBench, to test reconstructions under real physics. On the team's own numbers, scene-penetration errors drop from 69.51% to 22.90%, and interaction stability climbs to 53.68% on easy scenes and 30.56% on medium ones. On the hardest tier it reaches 13.92%, a result the team itself flags as the limit of today's approach.

Code and a project page are public. Optimized motions have been transferred to a Unitree G1 humanoid, showing the same sitting action runs on a real machine. Success rates on the physical robot were not disclosed.

No third-party reproduction has been published yet.