ASethi04/pi05-chris-traces-v1

Pi0.5 DROID Finetuned on Chris Traces v1

This model is a finetuned version of Physical Intelligence's Pi0.5 model, trained on the brandonyang/combined_chris_with_traces_v1 dataset.

Model Details

• Base Model: Pi0.5 DROID checkpoint
• Training Config: pi05_droid_finetune
• Training Steps: 20,000
• Batch Size: 32
• Training Time: ~17 hours on 8x H100 GPUs

Usage

This checkpoint is designed to be used with the openpi framework.

# Serve the policy
uv run scripts/serve_policy.py policy:checkpoint \
  --policy.config=pi05_droid_finetune \
  --policy.dir=ASethi04/pi05-chris-traces-v1

Training Metrics

Final training metrics at step 20,000:

• Loss: converged
• Logged to wandb: https://wandb.ai/rag-vla/openpi/runs/wkoh8r5t

Files

• params/ - Model parameters for inference
• assets/ - Normalization statistics (from original DROID checkpoint)
• _CHECKPOINT_METADATA - Checkpoint metadata