Vision-Language-Action (VLA) models have shown promise in translating multimodal context into robotic actions, but their action decoders are typically trained via behavior cloning, which supervises the demonstrated motor commands while leaving the underlying objective implicit. This oversight can limit the model's ability to generalize across varied situations.
To address this, researchers at Hugging Face propose Intention Distillation (INDI), a novel training approach that distills behavior-level intent into the action decoder. During training, a frozen teacher vision-language model (VLM) interprets a demonstrated segment—including the current observation, instruction, coarse action summary, and execution video—to recover a multimodal intent representation. At deployment, the VLA model uses this recovered intent at an intermediate decoder layer to organize action prediction alongside representations of how the behavior unfolds and what it achieves.
INDI demonstrates consistent improvements across multiple benchmarks. On SimplerEnv-Bridge, it boosts GR00T-N1.7 from 64.3% to 84.7%, and on RoboCasa Kitchen from 64.1% to 70.3%. It also yields gains on π₀.₅ across both benchmarks. In real-world tasks, INDI improves average success from 62.0% to 68.7%, with gains up to 12.0 percentage points on longer-horizon tasks.
The authors' analyses confirm that the recovered latent representation is actively used by the decoder, captures both the behavior objective and execution progress, and organizes downstream predictions in an objective-dependent manner. These findings suggest that explicitly modeling the semantic objective of generated behavior can significantly enhance action decoder performance.