π0.5 — Pick and Place Toy to Bucket
This repository contains epoch 3 and epoch 4 JAX/Orbax inference checkpoints from a full fine-tune of Physical Intelligence's π0.5 model on the pick_and_place_toy_to_bucket dataset.
Checkpoints
| Directory | Epoch | Local checkpoint label | Optimizer updates |
|---|
epoch-3/ | 3 | 2616 | 327 |
epoch-4/ | 4 | 3487 | 436 |
The local OpenPI trainer logs microsteps and uses zero-based loop indices for checkpoint labels, while the LingBot checkpoint uses optimizer updates. Gradient accumulation is 8, so the final local checkpoint was labeled 3487 after all 3488 microsteps and 436 optimizer updates completed.
Each uploaded checkpoint contains params/ and assets/. The optimizer train_state/ is intentionally omitted because these are inference weights, not resume-training archives. The assets include the quantile normalization statistics used during training.
Training setup
- Base model:
pi05_base
- Framework: Physical-Intelligence/openpi
- Fine-tuning: full-parameter, 4 GPUs with FSDP
- Episodes / frames: 82 / 27,962
- Action horizon: 50
- Per-GPU micro-batch size: 8
- Gradient accumulation: 8
- Effective global batch size: 256
- Optimizer updates per epoch: 109
- Learning rate: constant
2e-5
- Epochs: 4
The observations contain front and side RGB images plus a 6-dimensional robot state. Actions are 6-dimensional: the first five joint dimensions are trained as deltas and the gripper dimension is absolute.
Loading with OpenPI
Use the same OpenPI checkout containing the pi05_toy_bucket_full_finetune training configuration and toy-bucket policy transforms:
1from pathlib import Path
2
3from huggingface_hub import snapshot_download
4from openpi.policies import policy_config
5from openpi.training import config as train_config
6
7repo_dir = Path(snapshot_download("Aikwed/pi05-pick-and-place-toy-to-bucket"))
8checkpoint_dir = repo_dir / "epoch-4"
9
10config = train_config.get_config("pi05_toy_bucket_full_finetune")
11policy = policy_config.create_trained_policy(config, checkpoint_dir)
Use epoch-3 instead of epoch-4 to load the third-epoch checkpoint.
Intended use
These weights are research artifacts for the matching robot embodiment, observation mapping, and action convention. They have not been evaluated for deployment on other robots. Validate outputs in simulation or with appropriate physical safety constraints before operating hardware.
