LeWAM: A JEPA World Action Model with Diffusion-Steering-Based MPC
LeWAM ties action prediction, dynamics, policy, and planning into one decoder-free JEPA latent model.
The paper says its bidirectional transformer is trained end-to-end across forward, backward, inverse dynamics, and policy prediction. Its latent representation made robot and object state easier to read with linear probes than a forward-only JEPA world model, while still filtering visual distractors. In closed-loop tests, LeWAM matched a same-size flow-matching policy and also retained world-model capability. For MPC, the authors report better performance when planning in the policy head’s noise space instead of sampling raw actions. ArXiv · AI/CL/LG's note
The paper says its bidirectional transformer is trained end-to-end across forward, backward, inverse dynamics, and policy prediction. Its latent representation made robot and object state easier to read with linear probes than a forward-only JEPA world model, while still filtering visual distractors. In closed-loop tests, LeWAM matched a same-size flow-matching policy and also retained world-model capability. For MPC, the authors report better performance when planning in the policy head’s noise space instead of sampling raw actions. ArXiv · AI/CL/LG's note
score 5