Reproduce “BOSS: Benchmark for Observation Space Shift in Long-Horizon Task” 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
Yue Yang, Linfeng Zhao, Mingyu Ding, Gedas Bertasius, Daniel Szafir
A new benchmark called BOSS has been introduced to investigate Observation Space Shift (OSS) in visual-input long-horizon robotic tasks, revealing that current imitation learning methods suffer substantial performance degradation due to OSS and that data augmentation alone is insufficient to address it.
Key insights
- Observation Space Shift (OSS) is formally defined as visual changes in skill-irrelevant predicates that negatively impact visuomotor policies during sequential skill execution.
- Data augmentation, even with extensive and diverse datasets, proved largely insufficient to mitigate OSS for existing imitation learning baselines, suggesting a need for algorithmic solutions.
- The severity of Observation Space Shift (OSS) is directly correlated with the magnitude of visual modifications, but not with the inherent difficulty of the skill being executed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.