Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models
INDI teaches robot action decoders the intent behind a demonstrated behavior, not just the next motor command.
The paper says a frozen teacher VLM derives a multimodal intent representation from the current observation, instruction, coarse action summary, and execution video. The deployed VLA then learns to recover that intent inside its decoder while predicting actions. Reported gains include GR00T-N1.7 rising from 64.3% to 84.7% on SimplerEnv-Bridge and from 64.1% to 70.3% on RoboCasa Kitchen. Real-world task success increased from 62.0% to 68.7%, with larger gains on longer-horizon tasks. Source: HF Daily Papers' note
The paper says a frozen teacher VLM derives a multimodal intent representation from the current observation, instruction, coarse action summary, and execution video. The deployed VLA then learns to recover that intent inside its decoder while predicting actions. Reported gains include GR00T-N1.7 rising from 64.3% to 84.7% on SimplerEnv-Bridge and from 64.1% to 70.3% on RoboCasa Kitchen. Real-world task success increased from 62.0% to 68.7%, with larger gains on longer-horizon tasks. Source: HF Daily Papers' note
score 5