EMPIRIC agent learns residual world models robots can program for new environments
tomssilver · x · 2026-10-04
- Problem: agentic real-to-sim can rebuild scenes, but cannot infer how glue cures, water heats, or fan wind pushes objects—mechanisms simulators often can't model.
- Method: EMPIRIC (Yichao Liang, Kevin Ellis et al., MIT/Basis) learns a "residual world model": a physics engine extended with code for missing mechanisms.
- Results: solves all 25 runs across five simulated domains (coding-agent baselines solve 14–16); on a real robot it learns wind force and domino masses from just two gusts.
- Significance: a step toward robots that program their own physical model of novel environments.
Related event: EMPIRIC Agent Learns Residual World Models to Fill Physics Engine Gaps(2 posts)→
More from Embodied
- OpenHarness: open-source command center for coding agents hits 1.1k GitHub stars — dee_hw · 2026-10-05
- Wayve's unsupervised robotaxi predicted in 36 months, with London first — Sethwinterroth · 2026-10-05
- NERC 2026 robotics colloquium wraps; Brown University to host 2027 — tomssilver · 2026-10-04
- ElevenLabs opens Brussels office with panel on voice AI as a robot interface — lukas_m_ziegler · 2026-10-04
- Alexandr Wang mocks the tiny screen on a DIY Muse walkie-talkie built in 1 hour — alexandr_wang · 2026-10-04
- Daimon's Tactile World Model Threads Beads at IROS by Feel, Not Just Vision — CyberRobooo · 2026-10-04