Reproduce “LIMBO: Learning and Internalizing Model-Free Barrier Objectives for Agile and Safe Whole-Body Control” 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
Jake Gonzales, Arturo Flores Alvarez, Yu-Ming Chen, Aaron D. Ames, Lillian J. Ratliff, Manikantan Nambi
LIMBO learns an action-level safety score from black-box transitions and a supplied failure margin, then uses it to train policies over residual actions around a frozen walking controller. Its Q-CBF focuses learning near recoverability boundaries without a hand-crafted barrier or action override. In simulation, changing replay concentration shifted behavior from crouching toward backward-leaning, with clearance rates of 85.7% to 91.2%. Hardware tests showed similar body lowering but greater drift for limbo; results are empirical, not an unconditional safety guarantee.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.