aswinkumar99/LeRobot-SO101-ACT-universal-all_bs32_s60000

LeRobot SO101 ACT universal-all_bs32_s60000

Summary

This repository contains the final checkpoint for a ACT policy 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 ACT policy was trained for this dataset configuration and was not initialized from a published ACT base checkpoint. The visual backbone uses ImageNet-initialized ResNet-18 weights (ResNet18_Weights.IMAGENET1K_V1) as recorded in the training config.

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/act/universal-all_bs32_s60000

• Batch size: 32

• Training steps: 60000

• Checkpoint save frequency: 15000

• 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: 2026-04-24T06:40:37.855245+00:00

• Training end: 2026-04-24T09:49:35

• Approximate training duration: 3h 8m 57s

• Vision backbone: resnet18

• Observation cameras: observation.images.overhead, observation.images.wrist

• Action chunk size: 100

• Action steps predicted: 100

Exact Training Command

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/act/universal-all_bs32_s60000 \
  --job_name=act_universal-all_bs32 \
  --batch_size=32 \
  --steps=60000 \
  --log_freq=200 \
  --save_freq=15000 \
  --save_checkpoint=true \
  --num_workers=8 \
  --wandb.enable=true \
  --wandb.project=so101-layout-generalization \
  --wandb.mode=online \
  --wandb.disable_artifact=true \
  --policy.type=act \
  --policy.device=cuda \
  --policy.push_to_hub=false

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-ACT-universal-all_bs32_s60000