Reproduce “Learn from What We HAVE: History-Aware VErifier that Reasons about Past Interactions Online” 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
Yishu Li, Xinyi Mao, Ying Yuan, Kyutae Sim, Ben Eisner, David Held
Researchers from Carnegie Mellon University and Tsinghua University introduced History-Aware VErifier (HAVE), a system enabling robots to disambiguate uncertain manipulation scenarios online by learning from past interactions. This approach consistently improved success rates and reduced steps to completion across various tasks, including articulated objects, multimodal doors, and uneven object pick-up, compared to history-conditioned generative models.
Key insights
- A generation-verification paradigm can theoretically and empirically improve expected action quality, particularly in environments with heterogeneous optimal solutions that challenge purely generative models.
- Explicitly decoupling action generation from history-aware verification allows the generator to propose diverse actions while the verifier reasons about their outcomes, circumventing mode collapse issues.
- The use of dense action fields and explicit observation flow, combined with a specialized attention mechanism, enhances the verifier's ability to robustly learn from and apply interaction history.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.