alexhegit/so101-simstudio-lab01-pnp-fastwam

SO-101 Lab01 Pick-and-Place — FastWAM (transfer failure)

FastWAM fine-tuned on the Lab 01 50-episode leader dataset. This is a documented negative result for LIBERO→SO-101 joint-space transfer, not a classroom pick-and-place policy. Same data: ACT 6-D 58% and MolmoAct2 6-D 54% full-range; this checkpoint never approaches the cube (reach_8cm=0).

Training data: alexhegit/so101-simstudio-lab01-pnp

This Hub revision is the best measured Lab 01 FastWAM run (still 0% closed-loop): start from the LIBERO 2-cam checkpoint `ZibinDong/fastwam_libero_uncond_2cam224` (not bare lerobot/fastwam_base, which has no action pathway tensors), pad 6 SO-101 joints into the 7-D action / 8-D proprio slots, freeze the Wan video expert then unfreeze the last 4 video MoT layers for 8K steps. lambda_video=0 (action-only; not the paper’s video co-training). Checkpoint 008000. Open-loop RMSE 0.136 rad (~3.8× ACT). Do not prefer the earlier fastwam_base 6-D 4K/20K runs (worse open-loop, still 0%).

Cameras: camera_top → image, camera_wrist → image2 (key-name alignment only; Lab 01 top-down ≠ LIBERO agentview). Eval in SimStudio unpads the 6 real joints (LAB01_FASTWAM_PAD_ACTION). Failures are before approach: the arm moves, the cube stays at spawn z≈0.015. Visual probes still respond to cube pixels; the defect is cross-embodiment action/coordinate grounding, not a dead camera. Ledger: Lab 01 §6.4.2.

Closed-loop (do not treat as a working policy)

Full-range random spawn, reset_arm: home, sync, n_action_steps=32:

Earlier Lab 01 FastWAM recipes (fastwam_base freeze-video 4K/20K, pad7 without unfreeze-4) are also 0% and are not this Hub revision.

Documentation

Quick load

from lerobot.policies.fastwam.modeling_fastwam import FastWAMPolicy

policy = FastWAMPolicy.from_pretrained("alexhegit/so101-simstudio-lab01-pnp-fastwam")

Sim2sim eval: LAB01_EVAL_CONFIG=labs/lab01_pnp/configs/rollout_fastwam.yaml and the FastWAM pad eval patch in eval.cmd. Not a demo policy.