LeWAM: end-to-end JEPA world model plans 32x faster, lifts robot success to 89.7%
randall_balestr · x · 2026-10-10
- LeWAM (UCSD/ETH Zurich/UBC/Brown) is the first end-to-end JEPA that jointly trains world modeling and action via next-latent prediction plus action flow matching.
- Planning results: contact-rich manipulation success jumps 28.6% → 89.7%, long-horizon tasks 10.9% → 46.2%, with 32.3x faster planning (19ms per plan).
- Tiny footprint: 17M params, one GPU, one hyperparameter — and validated on a real robot, not just simulation, with behavior-clone results competitive with established methods.
- Paper, code, data and checkpoints are open-sourced; interpretability analyses show control-relevant representations guiding test-time planning.
More from Embodied
- Vision Pro in 2026: motion controllers, best avatars, ultrawide display — still the most complete headset — bradneuberg · 2026-10-10
- NVIDIA Research's LPR team is hiring PhD interns in embodied AI — jonLorraine9 · 2026-10-10
- PrismaX uses skilled teleoperators for robot training data, sells datasets without sales calls — jonstephens85 · 2026-10-10
- Tesla Cybercab fleet accelerates: 319 registered in Texas, 90 on the road — JOBhakdi · 2026-10-10
- WorkOS gave 500 devs an ESP32-S3 conference badge; they put DOOM on it — DynamicWebPaige · 2026-10-10
- Hand-built robotic skin: AI asks to plug it into its head, reads touch in a minute — repligate · 2026-10-10