Updated leaderboards on the benchmark

[cvg25]

Hello @seohongpark,
First of all, thanks for putting together such a great benchmark suite for offline goal-conditioned learning. I have a couple questions:

1. I have been working from some months now on a new method and seems like for most tasks it beats the leaderboards published in the project website. However, I'd like to know if these learderboards are updated or are there any new SOTA models. Although, I am still in a experimental phase, I'd like to put toghether a paper if results continue to be as good :)

2. Regarding one of the research questions in the project page:

How can we combine expressive policies with goal-conditioned RL? If we compare the results on cube-single-play and cube-single-noisy, we can see that current offline goal-conditioned RL methods often struggle with datasets collected by non-Markovian policies in manipulation environments. Handling non-Markovian trajectory data is indeed one of the major challenges in behavioral cloning, for which many recent behavioral cloning-based methods have been proposed (e.g., ACT and Diffusion Policy). Can we incorporate these recent advancements in behavioral cloning into offline goal-conditioned RL?

My questions is, what prevents Diffusion Policy from being directly applicable to offline goal-conditioned RL problems like ogbench?

Regards,
Carlos

[Henry0132]

Hi, Carlos,

Not a maintainer, but an enthusiastic follower of offline goal-conditioned RL — so I couldn’t resist sharing a quick answer 😊.

Actually, Diffusion Policy has been successfully adapted to ogbench-style tasks! Recent works like “Elastic-Subgoal Diffused Policy Learning” and “State-Covering Trajectory Stitching for Diffusion Planners” show strong results. Nothing fundamentally blocks its use — it’s more about how you condition the policy on goals and integrate it with offline RL objectives.

Check arXiv — there’s a growing body of work here. And congrats on beating the leaderboard — consider sharing your method when ready!

[cvg25]

Hey @Henry0132,
Thanks a lot for the guidance. I didn't know about the "Elastic Subgoal Diffusion Policy" and its great work tackling a key aspect in long horizons. I'd like to benchmark against both methods, however, implementing them and running experiments on all OGBench may be too much. Do you know if there are any available results already?

I'm still working on the method, once it is ready I'll definitely share it :)

[Henry0132]

Hey Carlos,

Glad the pointers were useful! 😊
I’ve only run quick sanity checks on a couple of datasets (e.g., pointmaze-medium-navigate-v0), and the results roughly match what’s reported in the papers (within statistical variance). But unfortunately, I don’t have full ogbench evaluation results to share — running all tasks is indeed quite heavy!

Your best bet for now is still the authors’ repos or watching the official leaderboard for future submissions. Looking forward to seeing your method there soon!

[cvg25]

Hey @Henry0132,
I am writing as a follow-up on the topic. I have released a preliminary paper on my method (PaD) that achieves 95% on cube-single-play-v0 and 98 on cube-single-noisy-v0. I'll also share the code publicly soon.

I would like to test the method against other tasks in the benchmark but I'm currently very constrained by compute... anyways, I thought you'd like to give it a check, and I am open to discuss further :)

Here is the link to arxiv: Planning as Descent: Goal-Conditioned Latent Trajectory Synthesis in Learned Energy Landscapes