La-JEPA: Huawei-led 1B robot policy beats 12.4B Fast-WAM (60.8% vs 39.9%) and runs 5x faster
burny_tech · x · 2026-11-11
Huawei, University of Toronto, McGill and Mila introduce La-JEPA, a compact world-action model trained from scratch in the feature space of a pretrained V-JEPA 2.1 encoder — predicting future latents and actions during training but only actions at inference.
- RoboTwin 2.0: 1B model hits 60.8% vs 39.9% for the 12.4B Fast-WAM, while running 5x faster
- SOTA among policies without embodied pretraining on RoboTwin 2.0 and LIBERO-Plus
- Real-robot success comparable to π0.5 across four tasks
- Design study: JEPA features beat video VAE latents, especially out of distribution; training-time future prediction matters even when skipped at inference
- Authors also correct corrupted LIBERO-Plus instructions from the official evaluator and re-evaluate baselines
Takeaway: robot foundation models may benefit from control-first architectures rather than ones inherited from generation.
More from Embodied
- Supersplat XR update: Quest 3 streamed scenes jump from 9.7 to 30+ fps, Vision Pro gets WebGPU — willeastcott · 2026-11-11
- Cartwheel's thesis: generative kinematics atop a 50Hz tracking policy, no teleop grind — andrew_n_carr · 2026-11-11
- Cartwheel releases three open motion-tracking policies for humanoid robots on Hugging Face — andrew_n_carr · 2026-11-11
- Claude Fable 5.2 designs a humanoid robot overnight with CAD, sim and reel — davidpattersonx · 2026-11-11
- Russ Tedrake's CoRL talk: deploy robots early, train on intervention data to scale — DominiqueCAPaul · 2026-11-11
- Gemini Robotics humanoid serves snacks live on CoRL 2025 show floor — ki_ki_ki1 · 2026-11-11