Reproduce “Imagining Goals with Geometric Horizon Models for Policy Generalization” 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
Daniel Lawson, Dhruv Sreenivas, Adriana Hugessen, Khimya Khetarpal, Glen Berseth
TD-Aug augments offline goal-conditioned behavior cloning with imagined future goals to help policies compose routes from separate demonstrations. A geometric horizon model estimates discounted future states for mixed dataset behaviors, whose samples supplement within-trajectory goals. Across pointmaze stitching tasks, success averaged 79% versus 31% for GCFBC, with 54% versus 0% on giant-stitch. These offline benchmark results do not establish real-robot performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.