BLOG ILLUSTRATION / EXECUTION STORY

FAST encoding and continuous generation

The two lanes show FAST demonstration encoding and continuous flow generation as independent interfaces. Knowledge Insulation can use the token objective to train the backbone. Its continuous expert controls execution.

FAST encoding and continuous generationThe FAST lane encodes demonstrated motion as DCT/BPE tokens. The flow-matching lane jointly generates a continuous chunk from noise. The lanes are independent action interfaces. The two lanes are not a necessary sequence of runtime stages, and no data edge between them is necessary. FAST · encode demonstrated motion Flow · generate continuous motion Action chunk Normalize each dimension Temporal DCT Quantize coefficients Sequence + BPE Low-frequency first Noise + context Candidate + conditions Velocity field Integrate jointly Continuous chunk Motor commands
FAST encodes demonstrated action chunks as compact frequency-structured tokens. Flow matching jointly generates a continuous chunk from noise. Knowledge Insulation can use the token objective to train the backbone. Its continuous expert supplies inference-time actions. The author made this explanatory diagram. The playback is an example, and animation time is not measured runtime. Open full-size SVG