act-so101-multicolor-early-data-100k

ACT SO-101 Multicolor Early Data 100K

This is an Action Chunking Transformer (ACT) policy trained from scratch for an SO-101 follower arm.

Training data

• Dataset: ted88168/so101_multicolor_master_v1
• 243 episodes
• 175,923 frames at 30 FPS
• Cameras: observation.images.front and observation.images.handeye
• Robot state: 6 dimensions
• Action: 6 dimensions

Training

• LeRobot: 0.6.2-compatible source
• Steps: 100,000
• Steps 0-30,000: batch size 64
• Steps 30,001-100,000: batch size 48
• ACT parameters: 51,597,190
• Action chunk size: 100
• Vision backbone: ResNet-18 initialized with ImageNet weights

Intended use

Use this checkpoint as a synchronous ACT policy with the same SO-101 calibration, camera placement, image resolution, and workspace geometry used during data collection.

The checkpoint includes the policy configuration and LeRobot preprocessing/postprocessing files required for inference.

Important limitation

ACT is not language-conditioned. Although the training dataset contains red, green, and blue task labels, changing a text instruction at inference time does not reliably select a requested color. Evaluate grasping and placement with one target block first. For deterministic color selection, train separate ACT policies per color or use a language-conditioned VLA policy.

Safety

This model controls physical hardware. Start with short rollouts, keep the robot workspace clear, keep physical power cutoff immediately available, and do not treat a GUI stop control as an emergency stop.