Model Card for act (sorting_ACT)

Action Chunking with Transformers (ACT) is an imitation-learning method that predicts short action chunks instead of single steps. It learns from teleoperated data and often achieves high success rates.

This policy has been trained and pushed to the Hub using LeRobot. See the full documentation at LeRobot Docs.

This repository is a re-trained copy of adrrrobo/sorting_ACT, using the same architecture and dataset. It was trained for 15,000 steps on a single NVIDIA A100 40GB GPU (final training loss ≈ 0.22).

• Base / reference model: adrrrobo/sorting_ACT
• Dataset: adrrrobo/Hackathon2_20260606_210708 (99 episodes, ~106k frames)
• Policy: ACT, ~51.6M parameters
• Training: 15,000 steps, batch size 8

---

How to Get Started with the Model

For a complete walkthrough, see the training guide. Below is the short version on how to train and run inference/eval:

Train from scratch

lerobot-train \
  --dataset.repo_id=adrrrobo/Hackathon2_20260606_210708 \
  --policy.type=act \
  --output_dir=outputs/train/sorting_act \
  --job_name=sorting_act \
  --policy.device=cuda \
  --steps=15000 \
  --batch_size=8 \
  --wandb.enable=false

Writes checkpoints to outputs/train/sorting_act/checkpoints/.

Evaluate the policy/run inference

lerobot-record \
  --robot.type=so100_follower \
  --dataset.repo_id=<hf_user>/eval_<dataset> \
  --policy.path=kennytsu/sorting_ACT \
  --episodes=10

Prefix the dataset repo with eval_ and supply --policy.path pointing to a local or hub checkpoint.

---

Model Details

• License: apache-2.0
• Trained on: NVIDIA A100 40GB, 15,000 steps
• Final training loss: ~0.22