IntentionVLA: Generalizable and Efficient Embodied Intention Reasoning for Human�Robot Interaction
Yandu Chen, Kefan Gu, Yuqing Wen, Yucheng Zhao, Tiancai Wang, Liqiang Nie
AI summary
Problem
Current Vision-Language-Action models lack reasoning-intensive pretraining and fail to interpret implicit human intentions, causing them to struggle with contextual understanding and accurate execution in complex, real-world interactions.
Approach
The method employs a two-stage curriculum training paradigm that first equips a VLM backbone with embodied intention and spatial reasoning capabilities, then distills these into compact reasoning cues to guide a diffusion-based action generator for fast inference.
Key results
- 18% higher success rate than π0 and 28% higher than ECoT under intention instructions
- Doubles baseline success rates on out-of-distribution tasks
- Enables zero-shot human-robot interaction with 40% success rate
- Automated pipeline for generating intention, spatial, and compact reasoning data
Why it matters
Provides a scalable, real-time framework for next-generation human-robot interaction systems that require accurate interpretation of ambiguous, intention-driven commands.
Abstract
Vision-Language-Action (VLA) models leverage pretrained vision-language models (VLMs) to couple percep- tion with robotic control, offering a promising path toward general-purpose embodied intelligence. However, current SOTA VLAs are primarily pretrained on multimodal tasks with limited relevance to embodied scenarios, and then finetuned to map explicit instructions to actions. Consequently, due to the lack of reasoning-intensive pretraining and reasoning- guided manipulation, these models are unable to perform implicit human intention reasoning required for complex, real- world interactions. To overcome these limitations, we propose IntentionVLA, a VLA framework with a curriculum training paradigm and an efficient inference mechanism. Our proposed method first leverages carefully designed reasoning data that combine intention inference, spatial grounding, and compact embodied reasoning, endowing the model with both reason- ing and perception capabilities. In the following finetuning stage, IntentionVLA employs the compact reasoning outputs as contextual guidance for action generation, enabling fast inference under indirect instructions. Experimental results show that IntentionVLA substantially outperforms π0, achieving 18% higher success rates with direct instructions and 28% higher than ECoT under intention instructions. On out-of-distribution intention tasks, IntentionVLA achieves over twice the success rate of all baselines, and further enables zero-shot human- robot interaction with 40% success rate. These results highlight IntentionVLA as a promising paradigm for next-generation human-robot interaction (HRI) systems.