INDI Lifts GR00T-N1.7 Success from 64.3% to 84.7% on SimplerEnv
TL;DR
- INDI raises GR00T-N1.7's SimplerEnv-Bridge success rate from 64.3% to 84.7% by distilling behavior-level intent into the action decoder.
- RoboCasa Kitchen improves from 64.1% to 70.3%, and real-world task average moves from 62.0% to 68.7% under the same method.
- A frozen teacher VLM supplies the semantic objective, and the paper reports gains reaching up to 12 percentage points on longer-horizon tasks.
GR00T-N1.7's success rate on the SimplerEnv-Bridge benchmark climbs from 64.3% to 84.7% when trained with a method the authors call Intention Distillation, or INDI, according to a new paper from Sangoh Lee, Sangwoo Mo, and Wook-Shin Han.
The setup: a frozen teacher VLM sits alongside the policy during training and is used to interpret what the demonstrator was trying to do, not just the motor sequence they produced. In the authors' phrasing, INDI "distills behavior-level intent into the action decoder," with the decoder recovering multimodal intent representations at an intermediate layer.
Gains elsewhere are smaller. On RoboCasa Kitchen, controlled GR00T-N1.7 moves from 64.1% to 70.3%. On a real-world suite, average success rises from 62.0% to 68.7%. The paper reports gains reaching up to 12 percentage points on longer-horizon tasks.
Originally reported by paper
Read the original article →Original headline: INDI Pushes Robot VLA Success Rate From 64% to 85% by Distilling Semantic Intent, Not Just Motor Commands