Pi0.5 Model - FR5 Garlic Pickup
This model is a fine-tuned Pi0.5 checkpoint trained on the FR5 garlic pickup dataset.
Model Details
- Base Model: Pi0.5 (Physical Intelligence)
- Task: Garlic pickup and placement using FR5 robotic arm
- Training Steps: 10,000
- Checkpoint: Step 10,000
- Config:
pi05_fr5_garlic_production - Experiment Name:
pickup_single_garlic
Training Configuration
- Model: Pi0.5 with action_horizon=50
- Batch Size: 32 (16 per GPU)
- FSDP Devices: 2 (H200 GPUs)
- Data: Custom FR5 garlic manipulation dataset (60Hz, ~22s episodes)
- Tolerance: 1.0s (for high-frequency data with large action horizon)
Model Files
This checkpoint contains:
params/: Model parameters in Orbax format (~12GB)train_state/: Training state including optimizer state (~30GB)assets/: Normalization statistics and other assets_CHECKPOINT_METADATA: Checkpoint metadata
Usage
This checkpoint can be loaded using OpenPI.
Training Details
- Framework: JAX
- Precision: bfloat16
- Training Time: ~10-12 hours on 2x H200 GPUs
- Dataset: FR5 robot manipulation data at 60Hz