Reproduce “IntentVLA: Short-Horizon Intent Modeling for Aliased Robot Manipulation” 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
Shijie Lian, Bin Yu, Xiaopeng Lin, Zhaolong Shen, Laurence Tianruo Yang, Yurun Jin, Haishan Liu, Changti Wu and 3 more
IntentVLA is a Vision-Language-Action framework that stabilizes robot manipulation by inferring short-horizon intent from recent visual history to resolve ambiguities in aliased observations. It achieved a 45.8% average success rate on the new AliasBench benchmark, significantly outperforming a frame-only baseline (9.0%), and improved success rates on various standard simulation and real-world tasks, including a 72.9% average on SimplerEnv.
Key insights
- Relying solely on the current observation frame is insufficient for VLA policies to maintain local commitment in partially observable scenarios with visually aliased states.
- Integrating recent visual history, even in a compact representation, effectively disambiguates visually similar states, leading to more stable and consistent short-horizon intent and action generation.
- Explicitly encoding 'short-horizon intent' through a combination of fine-grained history tokens and a compact history-evidence token significantly improves inter-chunk consistency and overall task success.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.