<ul><li>Offline goal-conditioned reinforcement learning is a promising approach for pretraining generalist policies on large datasets of reward-free trajectories.</li><li>Hierarchical RL methods achieve good results on long-horizon goal-reaching tasks but face challenges in scaling due to their reliance on modular policies and subgoal generation.</li><li>A new algorithm has been introduced to train a flat goal-conditioned policy by bootstrapping on subgoal-conditioned policies, eliminating the need for a generative model over the goal space.</li><li>This approach outperforms existing offline GCRL algorithms on various locomotion and manipulation benchmarks, scaling well to complex, long-horizon tasks.</li></ul>

Flattening Hierarchies with Policy Bootstrapping

Discover more