BLOG ILLUSTRATION / EXECUTION STORY

Condition a continuous action expert

Visual-language features and robot state condition continuous action generation. Generation updates a candidate before physical execution.

Condition a continuous action expertVLM features from camera input and instruction condition the flow-matching action expert. Current robot state also conditions the expert. The expert generates all positions of the continuous chunk before execution. Generation index s and robot time t are different. The diagram does not give the full loss or sampler. Condition continuous action generation Vision + language Cameras + instruction VLM backbone Semantic context Action expert Flow velocity field Robot state Proprioception q_t Action chunk Continuous commands Generation s ≠ robot time t
Visual-language representations condition a flow-matching action expert, with robot state as an input to the policy’s motor interface. The expert generates all positions of a continuous action chunk before robot execution. Generation time is different from robot execution time. The author made this explanatory diagram. The playback is an example. Animation time is not measured runtime. Open full-size SVG