alexhegit/so101-simstudio-lab01-pnp-vla-jepa

SO-101 Lab01 Pick-and-Place — VLA-JEPA

VLA-JEPA policy fine-tuned from lerobot/VLA-JEPA-LIBERO on expert demonstrations collected and validated with SO-101 SimStudio (MuJoCo sim2sim, leader-arm teleop).

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

This Hub revision is the MI300X run: batch 16, 20 000 steps (resume 10K→20K), world-model co-training on, chunk_size / n_action_steps 7, final train loss ~0.115. Wall time 3.5 h (first 10K) + **4.2 h** (10K→20K) on AMD Instinct MI300X (DORobot). Checkpoint 020000 / last.

Cameras: camera_top → image, camera_wrist → image2. Action is 6-D joint position. Proprio is Lab 01’s 15-D observation.state (6 pos + 6 vel + 3 EE); the LIBERO base config still lists 8-D input_features — load with state_dim=15 (see SimStudio eval.py).

Closed-loop MuJoCo (10K, fixed spawn, reset_arm: home, sync, GLFW): 0/10. The 20K weights on this Hub page have not been eval’d in sim yet.

Documentation

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

Quick load

from lerobot.policies.vla_jepa.modeling_vla_jepa import VLAJEPAPolicy

policy = VLAJEPAPolicy.from_pretrained("alexhegit/so101-simstudio-lab01-pnp-vla-jepa")