Action Chunking Transformer For The Koch v1.1 Robot

Action Chunking Transformer (ACT) trained on the AdamLucek/koch_block dataset for the Koch v1.1 Low Cost Robot from LeRobot

Usage

Details around methodology for set up, training, and evaluation available via LeRobot's guide to Getting Started with Real World Robots

Training Details

The model was trained using LeRobot's training script and the AdamLucek/koch_block dataset, using this command:

python lerobot/scripts/train.py `
  dataset_repo_id="AdamLucek/koch_block" `
  policy="act_koch_real" `
  env="koch_real" `
  hydra.run.dir="outputs/train/act_koch_test" `
  hydra.job.name="act_koch_test" `
  device="cuda" `
  wandb.enable="true"

Training was done on a 3060 for roughly 7 hours, loss curve and statistics available in wandb report here.