GEM-4D: geometry-enhanced video world models lift robot sim success 22.7% to 75.1%
pliang279 · x · 2026-11-13
A Harvard × MIT × Princeton team (ECCV 2026) presents GEM-4D, a geometry-grounded video world model for robot manipulation. It distills dense 4D correspondence supervision from a frozen geometry model into the video backbone during training, then drops that branch at inference — zero extra inference cost — yielding geometry-consistent rollouts an inverse-dynamics system reads into executable 6-DoF trajectories, closing the loop from one instruction and observation to a real arm with no task-specific training. Results: sim success on 7 tasks rises from 22.7% to 75.1% (+52.4pp, same action extractor); real-robot success improves 61% → 81%.
More from Embodied
- CoRL workshops overflow: attendees on floors and hallways at continual learning session — DominiqueCAPaul · 2026-11-13
- UW Prof. Abhishek Gupta muses on academia's role in robotics RL research at CSIR workshop — chris_j_paxton · 2026-11-13
- YC F26 startup Kairox builds autonomous drones that both fly and swim — ycombinator · 2026-11-13
- Mila brings its robotics research to CoRL 2026 in Austin — Mila_Quebec · 2026-11-13
- Fable 5.2 robot evals: overall success rate jumps from 34% to 50% across 456 tasks — GuanyaShi · 2026-11-13
- Day 8 of building a robot arm: maker ditches off-the-shelf gripper and builds his own — _Stocko_ · 2026-11-13