Reproduce “A Rose by Any Other Name Is Not as Sweet: How Rephrasings Induce Failures in Embodied AI” 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
Soumil Datta, Atharv Belsare, Ali Larian, Guanhong Tao, Daniel S. Brown
The authors test embodied robot planners on controlled rephrasings, comparing attempted actions with a hidden correct task description rather than final object positions alone. In MuJoCo simulation, 352 of 2,886 instruction-breaking plans passed the final-scene check. Qwen2.5-7B with JSON planning succeeded on 100% of forward instructions but 56% of matched reversed-order forms; this was not tested on physical robots.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.