Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models

TL;DR

INDI (Intention Distillation, Lee/Mo/Han) adds a training-time auxiliary objective to standard VLA behavior cloning: a frozen teacher VLM interprets a full demonstration segment (observation + instruction + coarse action summary + execution video) into a multimodal intent representation of the behavior’s semantic objective, and the action decoder is trained to recover that intent at an intermediate layer from its standard inputs alone (no video, no coarse summary at deployment). The recovered latent then organizes downstream action prediction. Reported gains on GR00T-N1.7 are +20.4 pp on SimplerEnv-Bridge (64.3 → 84.7%) and +6.2 pp on RoboCasa Kitchen (64.1 → 70.3%), with consistent gains on π₀.₅ across both benchmarks and real-world average success 62.0 → 68.7% (up to +12 pp on longer-horizon tasks). The teacher is discarded at inference.

Key claims

Standard behavior cloning supervises which motor command was demonstrated but leaves the local semantic objective of the behavior implicit; INDI proposes distilling that objective into an intermediate decoder layer as an explicit supervision signal [Abstract].
Future-based supervision (future frames, latent observations, trajectories, motion representations) captures particular realizations of what may happen rather than the shared semantic objective of the forthcoming behavior — INDI positions intent distillation as targeting a different level of abstraction than the future-prediction auxiliary-loss family [Abstract].
SimplerEnv-Bridge: GR00T-N1.7 improves from 64.3% to 84.7% with INDI (+20.4 pp) [Abstract].
RoboCasa Kitchen: controlled GR00T-N1.7 baseline improves from 64.1% to 70.3% with INDI (+6.2 pp); π₀.₅ shows consistent gains across both benchmarks [Abstract].
Real-world tasks: average success improves 62.0% → 68.7% with gains up to +12.0 pp on longer-horizon tasks — largest lift is at the horizon regime where implicit objective drift is most costly [Abstract].
Ablation-style analyses claim the recovered latent is actually used by the decoder, captures behavior objective and execution progress, and organizes downstream predictions in an objective-dependent manner [Abstract].

Method

Training: a frozen teacher VLM sees the full demonstration segment — current observation, natural-language instruction, coarse action summary, and the corresponding execution video — and produces a multimodal intent representation intended to encode what that behavior is trying to achieve. The deployed VLA sees only its standard inputs (observation + instruction, no video, no coarse summary) and is trained to recover that same intent representation at an intermediate layer of its action decoder. The recovered intent then serves as an internal organizing signal alongside auxiliary representations of how the behavior unfolds and what it achieves, on top of the standard action-prediction loss. Deployment: the teacher is discarded; the VLA infers intent from its own standard-input pathway and conditions action prediction on it.

The recipe is grafted onto existing VLA backbones (GR00T-N1.7 and π₀.₅ in the reported experiments) without changing the deployment interface — no new inputs, no runtime dependency on a teacher, no separate module invoked at inference.

Results

SimplerEnv-Bridge: GR00T-N1.7 64.3% → 84.7% with INDI, +20.4 pp [Abstract].
RoboCasa Kitchen: controlled GR00T-N1.7 baseline 64.1% → 70.3%, +6.2 pp; consistent gains on π₀.₅ across both benchmarks [Abstract].
Real world: 62.0% → 68.7% average success across tasks, with the largest per-task lifts (up to +12.0 pp) concentrated on longer-horizon tasks [Abstract].
The reported analyses claim the recovered latent (a) is actually used by the decoder (probing / causal ablation implied), (b) captures behavior objective and execution progress, and (c) organizes downstream predictions in an objective-dependent manner [Abstract].

Why it’s interesting

INDI sits on a distinct row of the VLA Models recipe-lever board from what’s currently filed: an auxiliary-loss lever that supervises the action decoder’s intermediate representation with an offline-computed intent latent, rather than adding modalities (T-Rex tactile), changing the value/reward pipeline (π*0.6 RECAP), or grafting a world model onto the perception side (µ₀, FullDiT2). The load-bearing move — teacher sees the full demonstration segment (including future video), student recovers the same latent from present-only inputs — is close in spirit to Anchor-Align: Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment Anchor-Align’s representation-anchoring loss but targets a behavior-level intent (a temporal-semantic abstraction over an executed segment) rather than a per-frame VLM representation.

The paper is also the closest sibling on file to LUCID: Learning Embodiment-Agnostic Intent Models from Unstructured Human Videos for Scalable Dexterous Robot Skill Acquisition LUCID on the axis of “what is intent, and where does it live in the pipeline.” LUCID retargets intent (short-horizon 3D object flow + palm pose) from human video and delegates control to a sim-trained sensorimotor policy; INDI distills intent from a teacher VLM interpreting robot demonstrations into the action decoder itself, and control is not delegated. If both approaches replicate, this contrasts two positions on intent-in-VLA design: (i) intent as an external interface that decouples pipeline halves (LUCID), (ii) intent as an internal representation-shaping signal that stays inside a single VLA (INDI). Worth flagging that INDI’s largest reported margins (+12 pp on longer-horizon real-world tasks) are exactly where implicit-objective drift under BC would be expected to hurt most, aligning with Behavioral Cloning Mystery‘s diagnosis of test-time distribution shift under Markovian BC.