LeWAM

LeWAM predicts 32 robot actions per inference call using 10 denoising steps by default. This repository contains the LeWAM predictor, the frozen I-JEPA-Huge vision encoder, the scheduler, and RoboTwin action statistics.

Download all weights together:

hf download XuejiFang/LeWAM --local-dir outputs/lewam

Use the LeWAM inference code for RoboTwin evaluation. The bundled I-JEPA-Huge encoder is licensed under CC BY-NC 4.0.