Robocasa_navigatekitchen

A pi0.5 (ฯ€โ‚€.โ‚…) Vision-Language-Action (VLA) model, finetuned on the ROBOCASA robotic manipulation/navigation benchmark using the OpenTau training framework. This model is designed to follow natural language instructions to perform manipulation/navigation tasks in a simulated kitchen environment.

For full documentation, evaluation results, and inference code, please visit the repository:
๐Ÿ‘‰ https://github.com/TensorAuto/OpenTau


Model Details

Description

  • Model Type: Vision-Language-Action (VLA) Model
  • Base Architecture: ฯ€โ‚€.โ‚… (pi0.5) by Physical Intelligence
  • Backbone: PaliGemma-3B (VLM) + Gemma-300M (Action Expert)
  • Training Data: Robocasa Benchmark
  • Framework: OpenTau

Architecture

The pi0.5 architecture uses a flow-matching-based policy designed for open-world generalization. It combines a Visual Language Model (VLM) for high-level semantic understanding with a smaller "action expert" model that generates continuous joint trajectories (10-step action chunks) via flow matching.


Training and Evaluation

Dataset

This model was finetuned on the Robocasa benchmark dataset. The Robocasa suite consists of human-teleoperated and mimicgen demonstrations for manipulation and navigation, covering:

  • CloseMicrowave (Atomic)
  • CloseFridge (Atomic)
  • CloseCabinet (Atomic)

Results

Training on 100 Human demonstrations, our model achieves 98% , 80% and 65% success rate on CloseMicrowave, Close Fridge and Close Cabinet tasks respectively. For detailed usage instructions, success rates, baseline comparisons, and evaluation protocols, please refer to the OpenTau GitHub Repository.

Downloads last month
23
Video Preview
loading