Find the branches.
Estimate multi-step predictive uncertainty and concentrate graph landmarks near high-variability, branching regions of offline trajectories.
NeurIPS 2026 / Offline Goal-Conditioned Reinforcement Learning
Escaping Path Mirages in Offline Goal-Conditioned Reinforcement Learning
Graduate School of Artificial Intelligence, UNIST
Not every promising path is reliable.
BGP plans around stochastic branches to reach distant goals more consistently.
01 / THE PROBLEM
When a successful path is a mirage.
In stochastic environments, apparently successful trajectories may depend on rare, lucky outcomes. Offline learners can mistake these paths for reliable routes, repeatedly entering branching regions that frequently lead to failure. We call this failure mode path mirage.
(a) Ambiguous downstream outcomes at a branching region. (b) Deceptive successes can mislead planning. (c) BGP identifies stochastic landmarks and selects reliable routes.
THE CHALLENGE
Connectivity is not reliability.
A transition that appears useful in recorded trajectories can still be risky when the environment has multiple stochastic outcomes.
OUR INSIGHT
Model where paths branch.
Identify high-variability states, then evaluate possible downstream outcomes before committing to a waypoint.
02 / OUR APPROACH
Branch-aware Graph Planning.
BGP builds a goal-conditioned planning graph that treats stochastic branching as a first-class part of long-horizon decision making.
Choose reliable routes.
Construct goal-conditioned edge costs that consider downstream branches. Compose local reachability estimates and plan waypoints with graph search.
Learn locally.
Train the policy and value function on trajectory windows bounded by landmark cluster transitions, reducing long-horizon error accumulation.
The result: reliable waypoint planning without treating all data-supported paths as equally trustworthy.
03 / PATH ANALYSIS
Seeing the difference in a teleport maze.
A teleport can lead to different locations, including unreachable dead zones. A local-value edge cost may favor an apparently short route through it; branch-aware edges instead account for downstream reachability and choose a route appropriate for the goal.
(a–b) Planning with the prior edge cost can enter deceptive teleport routes. (c–d) BGP's goal-conditioned edges account for downstream branches and yield successful waypoint paths. The same landmark set is used for comparison.
04 / EXPERIMENTS
Reliable goal reaching across OGBench.
BGP is evaluated on state-based and visual maze benchmarks, including challenging navigation, stitching, exploration, and stochastic teleport settings.
91%AntmazeMean success rate
98%PointmazeMean success rate
86%Visual-antmazeMean success rate
Mean success rates reported in Table 1 of the paper; each environment group aggregates its evaluated datasets.
Robust even as teleport stochasticity increases.
We vary the number of active teleport entrances (blackholes) from 0 to 4. Across navigate and stitch datasets, BGP remains substantially more successful than HIQL as teleport-induced stochasticity increases; segmented learning provides additional stability.
(a–b) Example mazes with two and four active teleports. (c–d) Success rates on navigate and stitch datasets as the number of teleports increases. Shaded regions show 95% confidence intervals across five random seeds.
| Antmaze teleport dataset | HIQL | GAS | BGP |
|---|---|---|---|
| Navigate | 42% | 39% | 90% |
| Stitch | 36% | 39% | 92% |
| Explore | 34% | 32% | 93% |
Selected comparison from Table 1. Success rates (%) averaged across five seeds; see the paper for standard deviations and all baselines.
05 / RESOURCES
Explore BGP.
Read the full paper, inspect the experiments, and explore the official implementation.
PaperOpenReview link coming soon
Source codeOfficial BGP GitHub repositoryUNIST MLICMore research projects from our lab06 / CITATION
Cite this work.
@inproceedings{bae2026escaping,
title={Escaping Path Mirages in Offline Goal-Conditioned Reinforcement Learning},
author={Bae, Junhyeon and Hwang, Jaebak and Choi, Gwanwoo and Kim, Minung and Han, Seungyul},
booktitle={Advances in Neural Information Processing Systems},
year={2026}
}