alexhegit/so101-simstudio-lab01-pnp-act-state6

SO-101 Lab01 Pick-and-Place — ACT (6-D joint state)

ACT policy fine-tuned on expert demonstrations from SO-101 SimStudio (MuJoCo, leader-arm teleop).

Training data: alexhegit/so101-simstudio-lab01-pnp (on-disk observation.state is still 15-D: 6 joint pos + 6 vel + 3 EE). This checkpoint was trained on the first 6 dims only (joint .pos), matching official real LeRobot SO-101 (so_follower) proprioception.

Why 6-D: so the policy input matches real-robot IL and is easier to sim2real / merge with real datasets. Extra sim channels (velocity, end-effector XYZ) are not available on the stock real follower. Units (radians vs degrees, gripper scale) still need a separate alignment.

The 15-D ACT reference (same 50K schedule, pos+vel+ee) is alexhegit/so101-simstudio-lab01-pnp-act.

This Hub revision

Item	Value
GPU	AMD Instinct MI300X (DORobot)
Batch / steps	128 / 50 000
Train loss	0.054
Sim2sim eval	full-range, reset_arm: follow, sync, EGL, n_action_steps=50 → 29/50 (58%)
15-D ACT same protocol	32/50 (64%) — same level at n=50

Wall time ~17 h. Checkpoint 050000 / last.

Documentation

Resource	Link
SimStudio repo	rocPAI-Forge/so101-simstudio
Lab 01 walkthrough	labs/lab01_pnp/lab01_pnp.md

Quick load

from lerobot.policies.act.modeling_act import ACTPolicy

policy = ACTPolicy.from_pretrained("alexhegit/so101-simstudio-lab01-pnp-act-state6")

hf download alexhegit/so101-simstudio-lab01-pnp-act-state6 \
  --local-dir ./outputs/hub/lab01_pnp_act_state6

Sim2sim eval: labs/lab01_pnp/configs/rollout_act.yaml (same YAML as 15-D ACT; rollout already sends joint .pos).