GR00T SO101 Ball In Box

Fine-tuned GR00T policy for the yeopeter1031/so101-ball-in-box LeRobot dataset.

Training

• Base model: nvidia/GR00T-N1.5-3B
• Dataset: yeopeter1031/so101-ball-in-box
• Policy type: groot
• Steps: 2500
• Batch size: 16
• Effective samples: 40000
• Image transforms: enabled
• Final checkpoint: 002500
• Final logged loss: 0.065

Image Transforms

This checkpoint was trained with LeRobot on-the-fly image transforms enabled:

• Color jitter: brightness, contrast, saturation, hue
• Sharpness jitter
• Random affine

The augmented frames are not materialized as a separate dataset; transforms are applied by the dataloader during training.

Files

• model.safetensors: fine-tuned policy weights
• config.json: policy config
• train_config.json: training config
• policy_preprocessor.json: input preprocessing config
• policy_postprocessor.json: output postprocessing config

Notes

This is an offline training checkpoint. Real robot or replay inference should be used to evaluate task success rate.