LeCun co-authored paper: inverse dynamics loss fixes JEPA world models dropping unstable modes
udmrzn · x · 2026-10-11
A new arXiv paper by Leonardo F. Toso, Yann LeCun, James Anderson and Oumayma Bounou shows that next-step prediction with anti-collapse regularization does not guarantee JEPA world models preserve controllable unstable modes — the training loss can be minimized while those modes are collapsed, making stabilization impossible from the learned representation.
The fix: add an action reconstruction objective (inverse dynamics loss) to encourage control-aware representations:
- Proven that exact action reconstruction makes the encoder injective on the finite-horizon reachable subspace, so no action-reachable state direction within H steps is discarded
- As H grows, the dominant eigenspace of the controllability Gramian converges to the controllable unstable subspace
- Results proven for linear systems and validated empirically
More from Embodied
- AgenticROS Cloud adds telepresence: turn a ROS 2 robot into your remote eyes, ears and AI agent — chrismatthieu · 2026-10-11
- Adding Flexi-Tac Touch Sensors to SO101 Robots Boosts Policy Convergence and Efficiency — k7agar · 2026-10-11
- Tesla Cybercab fleet hits 300+ vehicles in Austin, an 8x scale-up in one month — aelluswamy · 2026-10-11
- Qualcomm CEO predicts AI phone supercycle, smart glasses as top AI wearable — SuB8u · 2026-10-11
- Meta FAIR's RoboJEPA establishes first scaling law for multi-embodiment robot world models — ylecun · 2026-10-11
- WUJI Hand 2 nails pen-spinning dexterity via glove teleop and sim-to-real, open source — ChrisGPT · 2026-10-11