alexhegit/so101-simstudio-lab01-pnp-smolvla

SO-101 Lab01 Pick-and-Place — SmolVLA

SmolVLA policy fine-tuned from lerobot/smolvla_base 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 64, 50 000 steps (~3.2M sample updates), final train loss 0.018, ~26 GB of 192 GB HBM. Wall time ~7h 45m on AMD Instinct MI300X (DORobot).

A shorter Strix Halo / 8060S iGPU schedule (batch_size=4, 7500 steps) is documented in Lab 01; it is not this checkpoint.

Documentation

Resource	Link
SimStudio repo	rocPAI-Forge/so101-simstudio
Lab 01 walkthrough (record → train → eval)	labs/lab01_pnp/lab01_pnp.md

Camera keys in the dataset (camera_top / camera_front / camera_wrist) map to SmolVLA camera1 / camera2 / camera3. See Lab 01 §5 for rename_map and the MI300X training recipe (batch_size=64, steps=50000).

Quick load

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy

policy = SmolVLAPolicy.from_pretrained("alexhegit/so101-simstudio-lab01-pnp-smolvla")

Sim2sim eval in MuJoCo: follow §6 Policy eval in lab01_pnp.md (SmolVLA configs under labs/lab01_pnp/configs/rollout_smolvla*.yaml).