Summary

Robot imitation data is often multimodal: similar visual-language observations can be followed by different action chunks depending on the demonstrator’s short-horizon intent or task phase. IntentVLA is a history-conditioned VLA that encodes recent visual observations into a compact short-horizon intent representation to condition action chunk generation, preventing inter-chunk conflict and unstable execution caused by inconsistent intent resampling across replanning steps.

Key Contributions

  • History-conditioned VLA framework encoding compact short-horizon intent representations
  • AliasBench: a 12-task ambiguity-aware benchmark on RoboTwin2 that isolates short-horizon observation aliasing
  • Improves rollout stability and outperforms strong VLA baselines on AliasBench, SimplerEnv, LIBERO, and RoboCasa
  • Code available on GitHub

Significance

Observation aliasing is a pervasive but underexplored failure mode in VLA deployment; IntentVLA and AliasBench provide both a solution and a principled evaluation framework for this problem.