IliaLarchenko/behavior-1k-solution

Extends Physical Intelligence's Pi0.5 vision-language-action model with learnable task embeddings replacing language processing, correlated noise sampling in Flow Matching to capture action dependencies, and a "System 2" stage-prediction auxiliary task for non-Markovian reasoning. Training employs multi-step Flow Matching with 15 expert predictions per VLM step and delta action normalization, while inference uses soft action inpainting with correlation-aware guidance and cubic interpolation for 1.3x speedup, evaluated on the BEHAVIOR-1K robotic manipulation benchmark.

254 stars.