asgard-robot/pi05-s0101-right-arm-tape-pickup

π₀.₅ S0101 Right Arm Tape Pickup Model

This is a fine-tuned π₀.₅ model specifically trained for S0101 robot right arm tape pickup tasks.

Model Details

• Base Model: π₀.₅ (Physical Intelligence)
• Task: Right arm tape pickup manipulation
• Robot Platform: S0101 single arm robot
• Training Data: 35,275 samples from 30 episodes
• Training Steps: 735 (2 epochs)
• Batch Size: 96 (optimized for H200 GPU)
• Action Dimension: 32D (padded from 6D S0101 joints)
• Action Horizon: 12 steps

Training Configuration

• Learning Rate: 5e-5 (cosine decay)
• Optimizer: AdamW
• EMA Decay: 0.999
• Warmup Steps: 110
• Memory Usage: ~105GB GPU memory

Usage

Performance

• Training Loss: Final loss ~0.043
• Convergence: Stable training with good gradient norms
• Checkpoint Size: ~50GB (includes optimizer state)

Hardware Requirements

• Inference: >8GB GPU memory
• Training: >22.5GB GPU memory (LoRA fine-tuning)

Dataset

Trained on the ASGARD S0101 right arm tape pickup dataset:

• Episodes: 30
• Samples: 35,275
• Cameras: RealSense RGB + Wrist camera
• Actions: 6-DOF joint control + gripper

Citation

If you use this model, please cite the original π₀.₅ paper and mention this fine-tuned version.

License

Apache 2.0