openEuler/pi05

Model Card for PI0.5/BananaPick (IB-Robot)

PI0.5 policy (as per π0 and π0.5: Vision-Language-Action Flow Models for Robot Manipulation) fine-tuned for banana pick-and-place on a 1-arm SO-101 robot within the IB-Robot framework.

This checkpoint fine-tunes lerobot/pi05_base (PaliGemma 2B vision-language model + Gemma 300M action expert) on a teleoperated banana-pick dataset. The bundle ships PyTorch weights plus a BERT tokenizer for language-conditioned inference, and a single inference_manifest.json routing table (schema v3).

Repository Structure

├── config.json                                                 # PI0.5 策略配置
├── model.safetensors                                           # torch 权重 (~8.8 GB, bf16)
├── policy_preprocessor.json
├── policy_postprocessor.json
├── policy_preprocessor_step_2_normalizer_processor.safetensors
├── policy_postprocessor_step_0_unnormalizer_processor.safetensors
├── bert-base-uncased/                                          # 语言指令 tokenizer (5 files)
│   ├── tokenizer.json
│   ├── tokenizer.model
│   ├── tokenizer_config.json
│   ├── special_tokens_map.json
│   └── added_tokens.json
├── train_config.json                                           # 完整训练超参
└── inference_manifest.json                                     # 部署路由表 (schema v3, 权威)

The directory layout and every file path inside inference_manifest.json must stay in sync. Do not rename or relocate files — the manifest's sha256 checks and path bindings depend on them.

Deployment Backends

Read inference_manifest.json → deployments[<target>] to route to the right backend. This bundle ships PyTorch-only deployments (compiled Ascend/RKNN artifacts are published separately).

Target	Backend	Runtime	Artifact	Hardware
torch-cuda	torch	PyTorch	model.safetensors (in bundle)	NVIDIA GPU
torch-cpu	torch	PyTorch	model.safetensors (in bundle)	CPU

Input tensors: observation.state [6] float32, observation.current [6] float32, observation.images.top [3,480,640] NCHW, observation.images.wrist [3,480,640] NCHW. Output tensor: action [6] float32 (6-dim joint action: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper).

How to Get Started with the Model

See the IB-Robot project (particularly the inference_service) for instructions on how to load and deploy this model with ROS 2.

To load the PyTorch backend directly in Python:

from lerobot.common.policies.pi05.modeling_pi05 import PI05Policy

policy = PI05Policy.from_pretrained("openEuler/pi05")

For ROS 2 deployment, consume inference_manifest.json and route to deployments["torch-cuda"] or deployments["torch-cpu"] via the IB-Robot inference_service.

Training Details

• Policy: PI0.5 (PaliGemma 2B + Gemma 300M action expert)
• Base model: lerobot/pi05_base
• Robot: 1-arm SO-101
• Task: Banana pick-and-place
• Cameras: top, wrist (480×640, resized to 224×224 internally)
• Action dim: 6 (5 joints + gripper)
• Chunk size: 50 action steps, 50 executed per step
• Inference steps: 10 flow-matching denoising steps
• Dtype: bfloat16
• Tokenizer: bert-base-uncased (max_length=200)
• Normalization: VISUAL=IDENTITY, STATE=QUANTILES, ACTION=QUANTILES
• Gradient checkpointing: enabled
• Optimizer: AdamW (lr=2.5e-5, weight_decay=0.01, betas=[0.9, 0.95])
• Scheduler: cosine decay with warmup (1000 warmup, 30000 decay, decay_lr=2.5e-6)
• Batch size: 8
• Training steps: 100,000
• Seed: 1000

Model Architecture

PI0.5 is a Vision-Language-Action (VLA) flow model:

• Vision encoder: PaliGemma (SigLIP-based, 224×224 input)
• Language model: Gemma 2B (PaliGemma backbone, processes text + image tokens)
• Action expert: Gemma 300M (separate decoder for flow-matching action generation)
• Flow matching: 10 denoising steps to generate action chunks
• Parameters: ~4.1B (533M F32 + 3.6B BF16)

Citation

@software{ib_robot,
  title = {IB-Robot: Intelligence Boom Robot},
  url = {https://gitcode.com/openeuler/IB_Robot},
  license = {Apache-2.0}
}

@article{pi05,
  title = {π0 and π0.5: Vision-Language-Action Flow Models for Robot Manipulation},
  url = {https://arxiv.org/abs/2410.24132}
}