LIT — Latent Interface Training checkpoints

Stage-1 and Stage-2 checkpoints for Breaking the Vision–Action Shortcut: Latent Interface Training for Generalizable Robot Foundation Models. Code, reproduction steps and the project page: https://github.com/jianmanlincjx/LIT

Directory	Backbone	What it is
molmoact2/lit_stage2	MolmoAct2	final LIT model — Table I/II LIT; Table III last row
pi05/lit_stage2	π0.5	final LIT model — Table I/II LIT
fastwam/lit_stage2	FAST-WAM	final LIT model — Table I/II LIT
imagewam/lit_stage2	ImageWAM (FLUX.2 Klein 4B)	final LIT model — Table I/II LIT
*/lit_stage1	same four backbones	Stage-1 action prior (no images; SE(3)-conditioned) that Stage 2 starts from — same file layout as the Stage 2 of that framework

molmoact2/* and pi05/* are LeRobot policy directories — pass the directory to lerobot_eval --policy.path. fastwam/* and imagewam/* hold model.pt + config.yaml + dataset_stats.json — pass ckpt=<dir>/model.pt and EVALUATION.dataset_stats_path=<dir>/dataset_stats.json. Baselines: MolmoAct2 and π0.5 baselines were fine-tuned by us on LIBERO (available on request); FAST-WAM and ImageWAM baselines are the authors' released weights.

Real-robot demonstrations used in the paper (YAM bimanual, three tasks, LeRobot v3.0): https://huggingface.co/datasets/chinchinati/yam_bimanual_manipulation

Evaluate LIBERO-Plus with LIBERO_PLUS_FIX_LANG=1; Overall is the mean over the seven perturbation axes.