Reproduce “IntentionVLA: Generalizable and Efficient Embodied Intention Reasoning for Human-Robot Interaction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yandu Chen, Kefan Gu, Yuqing Wen, Yucheng Zhao, Tiancai Wang, Liqiang Nie
IntentionVLA is a Vision-Language-Action (VLA) framework designed for efficient embodied intention reasoning in human-robot interaction. It achieves an average 45% success rate in interpreting implicit human intentions from ambiguous instructions, significantly surpassing baselines (e.g., ECoT at 16.7%) while maintaining real-time inference at 0.72 seconds.
Key insights
- Explicitly training models to infer latent human intentions from ambiguous instructions, beyond explicit commands, is crucial for natural and effective human-robot interaction.
- Distilling complex reasoning into concise, actionable textual cues ('compact short reasoning') enables efficient, real-time inference (0.72s) without compromising reasoning capabilities.
- A curriculum-based training approach, decoupling the VLM's reasoning from the action generator's execution via learnable queries, significantly enhances both generalizability and action stability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.