OGBench: Benchmarking Offline Goal-Conditioned RL

A benchmark built specifically to test offline goal-conditioned reinforcement learning algorithms, with 8 environment types and 85 datasets.

OGBench was proposed by Seohong Park, Sergey Levine, and colleagues, and published at ICLR 2025. Goal-conditioned reinforcement learning trains a policy to reach any specified goal state; offline means training uses only a fixed dataset, with no further interaction with the environment. The authors argued that on older benchmarks, different algorithms scored too close together to tell apart, so they designed 8 categories of environments — including maze navigation, block manipulation, and puzzles — with 85 datasets that separately test trajectory stitching (combining pieces of incomplete trajectories into a new route), long-horizon reasoning, image observations, and environment stochasticity, plus another 410 standard offline-RL tasks. It is built on MuJoCo and the Gymnasium interface, installs with pip, and ships reference implementations of six algorithms including GCBC, GCIQL, and HIQL.

ExampleIn the antmaze-large-navigate-v0 dataset, a four-legged ant robot must walk from a starting point to a goal location specified at evaluation time inside a large maze; most environments provide 5 evaluation goals each.

Also called

OGBench

Related

Goal-Conditioned Reinforcement Learning · Offline Reinforcement Learning · D4RL · Benchmark · Implicit Q-Learning · MuJoCo (Multi-Joint dynamics with Contact)

Sources

OGBench: Benchmarking Offline Goal-Conditioned RL (ICLR 2025)
OGBench project page
GitHub: seohongpark/ogbench