airoa-org/airoa-pi05-hsr-base

pi05_hsr_base

pi05_hsr_base is a continual-pre-training checkpoint derived from the public pi0.5 vision-language-action (VLA) model for downstream mobile manipulation with the Toyota Human Support Robot (HSR). It is trained with AIRoA-formatted, dual-camera teleoperation data and is intended to serve as an initialization for task-specific real-robot adaptation.

Model summary

Property	Value
Base checkpoint	Public pi0.5 checkpoint
Model type	Vision-language-action policy with flow matching
Target embodiment	Toyota Human Support Robot
Continual-pre-training data	AIRoA 75-task training collection
Continual-pre-training duration	Approximately 0.8 epoch
Sampling	Task-uniform, then episode sampling within each task
Global batch size	2,048
Training hardware	64 NVIDIA H200 GPUs
Trainable components	Action expert; LoRA adapters on separate head- and hand-camera vision encoders
Frozen component	Language-model backbone
Model inputs used	Head RGB, hand RGB, proprioception, task instruction
Intended role	Initialization for downstream task-specific fine-tuning

Architecture and training

pi05_hsr_base starts from the public pi0.5 checkpoint and retains its flow-matching training objective. During continual pre-training:

• the language-model backbone is frozen;
• the action expert is updated;
• separate vision encoders for the head and hand camera streams receive LoRA adapters; and
• tasks are sampled uniformly before sampling episodes within each task.

Task-uniform sampling prevents the long-tailed frequency distribution of the corpus from making the largest task families dominate optimization. The reported checkpoint was trained for approximately 0.8 epoch with a global batch size of 2,048 on 64 NVIDIA H200 GPUs.

The model consumes RGB observations and robot proprioception.

Training data

The checkpoint is continually pre-trained on a 75-task training collection drawn from the AIRoA mobile-manipulation collection effort, packaged in the AIRoA dual-camera teleoperation schema (head RGB, hand RGB, and proprioception). This training data is publicly released as AIRoA MoMa 5k.

Inputs and outputs

Inputs used during training

• natural-language task instruction;
• head-camera RGB observation;
• hand-camera RGB observation; and
• robot proprioceptive state.

Output

The checkpoint predicts model-native robot actions through the pi0.5 action expert.

License

pi05_hsr_base is released under the Gemma Terms of Use. Use, redistribution, and derivative works of these weights are subject to those terms and to the Gemma Prohibited Use Policy.

Citation

If you use pi05_hsr_base, please cite both the AIRoA MoMa dataset and pi0.5.

Training dataset: airoa-org/airoa-moma-5k

@article{takanami2025airoamoma,
  author  = {Takanami, Ryosuke and Khrapchenkov, Petr and Morikuni, Shu and others},
  title   = {{AIRoA MoMa} Dataset: A Large-Scale Hierarchical Dataset for Mobile Manipulation},
  journal = {arXiv preprint arXiv:2509.25032},
  year    = {2025}
}

@article{physicalintelligence2025pi05,
  author  = {{Physical Intelligence} and Black, Kevin and others},
  title   = {{$\pi_{0.5}$}: A Vision-Language-Action Model with Open-World Generalization},
  journal = {arXiv preprint arXiv:2504.16054},
  year    = {2025}
}