Reproduce “Right Goal, Wrong Body: Re-grounding Frozen VLAs under Actuation Shift” 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
Umer Siddique, Peilang Li, Yongcan Cao
ReGround adapts a frozen VLA to changed robot actuation by inferring intended motion from action chunks, learning deployed-body dynamics through task-agnostic babbling, and adjusting execution after detecting mismatch. In simulated LIBERO tests with unchanged visual observations, π₀.₅’s test interquartile mean (IQM) success across eight tasks at 400 ms latency rose from 0.16 to 0.90, near the 0.91 oracle; physical-robot recovery remains untested.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.