This repository contains the final checkpoint for a SmolVLA policy fine-tune trained on aswinkumar99/universal-all for SO101 sponge pick-and-place experiments.

Dataset meaning: Universal combined dataset (all available layouts/tasks in this training matrix).

This SmolVLA policy is a fine-tune of lerobot/smolvla_base, as recorded by both the launch command (--policy.path=lerobot/smolvla_base) and the saved training config (pretrained_path: lerobot/smolvla_base).

Training Setup

Dataset repo: aswinkumar99/universal-all
Local dataset root during training: /home/riftuser/datasets_combined/aswinkumar99/universal-all
Output directory during training: /home/riftuser/outputs_matrix/smolvla/universal-all_bs128_s20000
Batch size: 128
Training steps: 20000
Checkpoint save frequency: 5000
Data loader workers: 8
WandB project: so101-layout-generalization
GPU: NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition
Python: CPython 3.12.13
CUDA: 12.9
Training start: 2026-04-24T07:22:33.434013+00:00
Training end: 2026-04-24T11:54:14
Approximate training duration: 4h 31m 40s
Base model: lerobot/smolvla_base
Observation camera rename map: {"observation.images.overhead": "observation.images.camera1", "observation.images.wrist": "observation.images.camera2"}
Action chunk size: 50
Action steps predicted: 50

Exact Training Command

bash

lerobot-train \
  --dataset.repo_id=aswinkumar99/universal-all \
  --dataset.root=/home/riftuser/datasets_combined/aswinkumar99/universal-all \
  --dataset.video_backend=torchcodec \
  --output_dir=/home/riftuser/outputs_matrix/smolvla/universal-all_bs128_s20000 \
  --job_name=smolvla_universal-all_bs128 \
  --batch_size=128 \
  --steps=20000 \
  --log_freq=200 \
  --save_freq=5000 \
  --save_checkpoint=true \
  --num_workers=8 \
  --wandb.enable=true \
  --wandb.project=so101-layout-generalization \
  --wandb.mode=online \
  --wandb.disable_artifact=true \
  --policy.path=lerobot/smolvla_base \
  --policy.device=cuda \
  --policy.push_to_hub=false \
  --rename_map={"observation.images.overhead": "observation.images.camera1", "observation.images.wrist": "observation.images.camera2"}

Repository Contents

pretrained_model/: final downloadable model artifacts for inference/loading
training_state/: optimizer, RNG, scheduler/state, and step information for resuming or auditability

Notes

This repo stores the final checkpoint that was uploaded from the cloud training workspace.
The checkpoint was trained with LeRobot tooling via lerobot-train.
For SO101 experiments in this workspace, the dataset source was created by Aswinkumar.

Creator

Aswinkumar

Website: aswinkumar.me
Hugging Face repo: https://huggingface.co/aswinkumar99/LeRobot-SO101-SmolVLA-universal-all_bs128_s20000

aswinkumar99/LeRobot-SO101-SmolVLA-universal-all_bs128_s20000

LeRobot SO101 SmolVLA universal-all_bs128_s20000

Summary

Training Setup

Exact Training Command

Repository Contents

Notes

Creator

aswinkumar99/
LeRobot-SO101-SmolVLA-universal-all_bs128_s20000