Reproduce “Learning Goal-Reaching Quasimetric Geometry From Finite-Time Reachability” 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
Daisuke Yamada, Travis Pence, Vikas Singh
ReQRL learns goal-reaching costs for offline reinforcement learning by supervising a quasimetric critic with finite-horizon reachable motions and estimated environment boundaries. It generally outperformed QRL and EikHiQRL on maze benchmarks, including 89% success on humanoidmaze-large-stitch and gains on carmaze tasks where sequences of controls enable movements unavailable instantaneously.
Key insights
- Finite-horizon reachability can provide local critic supervision that reflects composed controls, rather than only the directions available at an instant.
- Separating interior dynamics constraints from boundary constraints offers a practical approximation to the effects of both control limits and environmental obstacles.
- Performance depends on the quality of dynamics and boundary estimates, which in turn rely on suitable offline-data coverage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.