Behavior-Skill VLA Checkpoints
This repository provides the π0.5 checkpoints used in Behavior-Skill, a fine-grained benchmark for evaluating Vision-Language-Action (VLA) policies in long-horizon mobile manipulation tasks.
The released checkpoints cover the complete 50-task setting and a representative 12-task subset under both task-level and skill-level language conditioning.
Checkpoints
| Checkpoint | Training Scope | Language Condition |
|---|---|---|
pi05-pt50-task |
50 tasks | Task instruction |
pi05-pt50-skill |
50 tasks | Skill instruction |
pi05-pt12-task |
12 tasks | Task instruction |
pi05-pt12-skill |
12 tasks | Skill instruction |
All checkpoints are initialized from the official pretrained π0.5 checkpoint.
Task vs. Skill
The Task and Skill variants use the same π0.5 architecture and robot demonstrations within each experimental setting.
- Task: conditioned on the original task-level instruction.
- Skill: conditioned on the instruction of the current constituent skill.
Apart from the language condition, the two variants within the same setting share the same model architecture and training configuration.
Usage
For environment setup, checkpoint loading, training, and evaluation instructions, please refer to the Behavior-Skill code repository:
GitHub: https://github.com/mafangniu/Behavior-Skill
Acknowledgements
Behavior-Skill is built upon BEHAVIOR-1K and uses π0.5 as one of its representative VLA baselines.
We thank the authors of BEHAVIOR-1K, OmniGibson, and π0.5 for releasing their datasets, simulation environments, models, and code.
License
These checkpoints are fine-tuned from the official pretrained π0.5 model.
Use and redistribution of the released checkpoints are subject to the applicable licenses and terms of the original π0.5 model and upstream resources.