Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models
The paper reports sizable robot-task gains from training action decoders to recover a latent “intent” signal, not just copy demonstrated actions.
INDI uses a frozen teacher VLM during training to interpret a demonstrated segment from observations, instruction, coarse action summary, and execution video. The deployed VLA then learns to recover that multimodal intent representation inside its decoder while predicting actions. On SimplerEnv-Bridge, it raises GR00T-N1.7 from 64.3% to 84.7%; in real-world tasks, average success rises from 62.0% to 68.7%. The authors say analysis shows the latent captures objective and execution progress and shapes downstream predictions. ArXiv · AI/CL/LG's note
INDI uses a frozen teacher VLM during training to interpret a demonstrated segment from observations, instruction, coarse action summary, and execution video. The deployed VLA then learns to recover that multimodal intent representation inside its decoder while predicting actions. On SimplerEnv-Bridge, it raises GR00T-N1.7 from 64.3% to 84.7%; in real-world tasks, average success rises from 62.0% to 68.7%. The authors say analysis shows the latent captures objective and execution progress and shapes downstream predictions. ArXiv · AI/CL/LG's note
score 5