GAS constructs a graph in a learned temporal-distance representation and uses shortest-path search to select subgoals, outperforming prior offline HRL methods on stitching and exploration benchmarks.
This substantial reduction in the number of nodes significantly improves overall efficiency in task planning and execution (Algorithm 1)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Graph-Assisted Stitching for Offline Hierarchical Reinforcement Learning
GAS constructs a graph in a learned temporal-distance representation and uses shortest-path search to select subgoals, outperforming prior offline HRL methods on stitching and exploration benchmarks.