LeWAM
Joint video-action flow-matching model for robot control.
- Architecture: 12-layer DiT, dim 512
- Context frames: 32 @ 5 fps
- Future frames: 8
- Action dim: 6 @ 30 fps
- Cameras: image1, image2
Usage
With lerobot:
pip install "lerobot_policy_lewam @ git+https://github.com/ErykHalicki/lerobot_policy_lewam.git"
lerobot-record --policy.type lewam --policy.pretrained_path ehalicki/lewam-so101-pretrained
- Downloads last month
- 88