Reproduce “Task adaptation of Vision-Language-Action model: 1st Place Solution for the 2025 BEHAVIOR Challenge” 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
Ilia Larchenko, Gleb Zarin, Akash Karnatak
The authors adapt π0.5 for household tasks in the 2025 BEHAVIOR Challenge. They use correlated noise derived from demonstrations and carry the final four actions of each prediction into the next, correcting the new sequence during denoising. The system ranked first on public and private leaderboards, but evaluation used new spatial configurations of the same 50 tasks and did not isolate component contributions.
Key insights
- Learnable mixed-layer attention enables the action expert to dynamically integrate features from various VLM layers, optimizing information flow without fixed architectural choices.
- Generating correlated noise for flow matching, based on empirical action covariance, significantly improves training efficiency and consistency of the denoising process.
- Multi-task training over diverse tasks can lead to emergent recovery behaviors, suchs as picking up fallen objects, even without explicit recovery demonstrations.
- Simple, targeted heuristic correction rules at inference time, like the gripper correction, can dramatically improve task success rates for specific failure modes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.