README.md

ACT Model for Dice Manipulation with Bounding Box Data

This model was trained on the so100_dice_red_v2 dataset with enhanced state vectors that incorporate bounding box information from PaLI-Gemma object detection.

Training Details

• Base dataset: so100_dice_red_v2
• State vectors: Modified to include subtle guidance based on dice position
• Training steps: 50,000
• Final loss: 0.046

Usage

This model can be used with the LeRobot framework for picking up dice with improved accuracy.