Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models
Sangoh Lee, Sangwoo Mo, Wook-Shin Han
cs.RO, cs.AI, cs.CV
2026-08-25
INDI distills teacher-VLM behavior intent into the VLA action decoder. GR00T-N1.7 rises 64.3% to 84.7% on SimplerEnv-Bridge; real-robot success 62.0% to 68.7%.
Vision-Language-Action models map images and instructions to continuous motor commands, yet the action decoder is still trained mostly by behavior cloning. That loss says which command was demonstrated. It does not say what local objective that behavior serves under the instruction. The same reach-grasp-place sequence can mean stack-the-cube or drop-in-the-basket. The same objective can also be realized by very different trajectories.
Future frames, subgoal images, trajectories, and flow tell the decoder what a rollout may look like. Those signals supervise one realization, not the semantic objective shared across realizations. INDI's claim is narrower: the decoder should recover what the forthcoming behavior is for.
During training a frozen teacher VLM, Cosmos-Reason2-8B by default, reads the current observation, the instruction, a coarse action summary, and the demonstration video, then writes a functional-purpose statement. The intent target is a pooled middle-layer multimodal state from that pass. Textual purpose uses the final-layer states of the generated sentence. Visual outcome is a frozen encoding of the endpoint frame.
The deployed VLA still takes the usual inputs: image, instruction, proprioception. Learnable intent queries recover the teacher target at about half decoder depth. Later layers jointly predict the action with the original flow-matching loss, plus latent visual and textual groundings. Attention is asymmetric: intent queries do not see noised action rows, while action rows can attend to intent and groundings. On half of the training steps, non-intent rows lose direct access to VLM context, so information has to pass through intent. A mismatch loss requires matched intent to beat a batch-shifted intent on downstream loss by a margin.
At deployment the teacher, cached targets, and alignment heads are gone. GR00T-N1.7 grows from 3.46B to 3.50B parameters and from 56.1 ms to 61.5 ms per query.
On four SimplerEnv-Bridge tabletop tasks, INDI lifts GR00T-N1.7 from 64.3% to 84.7% (+20.4). EP-Basket jumps from 36.0% to 96.0%; the other three tasks average +7.1. A future-supervision variant stops at 68.0%. The same recipe moves π0.5 from 52.3% to 58.8%.
On RoboCasa Kitchen, 24 tasks with 100 demonstrations each, GR00T-N1.7 goes from 64.1% to 70.3%, 0.5 points under the published 3,000-demo checkpoint at 70.8%. π0.5 goes from 34.9% to 41.4%, including +14.0 on the Others group.
| Setting | Baseline | +INDI |
| Bridge / GR00T-N1.7 | 64.3% | 84.7% |
| Bridge / π0.5 | 52.3% | 58.8% |
| RoboCasa / GR00T-N1.7 (100 demos) | 64.1% | 70.3% |
| Real-world, three-condition mean | 62.0% | 68.7% |
Real-world evaluation uses four tabletop tasks and 50 trials per condition: 71.0% to 76.0% in-distribution, 62.0% to 68.0% with held-out objects, 53.0% to 62.0% with distractors. Longer tasks move more. Drawer Storage rises from 24.0% to 36.0% on the clean split.
Controls pin the gain on intent. On a shared Bridge eval run, groundings-only and a free latent both fall below the action-only baseline. Intent alone reaches 76.0%; full INDI reaches 85.5%. Swapping in another task's intent drops success from 84.5% on-diagonal to 45.2% off-diagonal. Forcing early or late phase content collapses success to 7.3% and 5.5%. Teacher choice is not free: Qwen3.5-9B as teacher scores 54.5%, under the 61.5% no-distillation run.
For teams already training GR00T or π0.5, this is extra supervision at train time and almost no extra module at inference. Predicting another future frame is the weaker medicine: +3.7 on Bridge versus +20.4 from intent. Stage traces on the real robot match that story. Cross-Bin Stacking is tied through fetching the base cube; the gap opens when the policy must switch to the second cube. Multi-stage work is where an intermediate objective state earns its keep.
This remains a decoder-side increment, not a new VLA backbone. Teacher quality and whether the task is long enough will decide whether those twenty points show up.
There is no standalone limitations section. The Bridge headline is carried by EP-Basket's +60 points; the other three tasks are +7.1. Short real-world tasks sit near ceiling, and Basket Nesting even slips from 94.0% to 92.0% on the clean split. The eval surface is four simulated tasks plus four tabletop skills, with no assembly or mobile manipulation. The teacher must watch demonstration video, and a weak teacher hurts the student. A free latent underperforms, so extra capacity is not the explanation, but mismatch loss, context dropout, and asymmetric attention are bundled. The paper does not isolate each mechanism on the real robot.