Research map / Reinforcement learning / Offline RL
Score Models for Offline Goal-Conditioned Reinforcement Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SMORe, a discriminator-free offline goal-conditioned RL method that matches mixture goal-transition distributions and learns a score function and policy directly, shown in a rebuttal to outperform GoFAR.
Problem
Offline goal-conditioned RL from sparse rewards: existing distribution-matching approaches such as GoFAR rely on a learned discriminator and a loose lower bound, which is brittle and unstable.
Approach
Formulates distribution matching over goal-transition (state-action-goal) distributions mixed with offline data, giving a discriminator-free objective that tightly relates to the GCRL objective and is optimized via in-sample maximization with expectile/AWR-style policy extraction.
Key results
- Rebuttal reports SMORe outperforms GoFAR across environments, with improvements in pick-and-place tasks where other baselines struggle.
- GoFAR with a reward instead of a discriminator performs poorly, reducing to sparse-reward RL (Table 13 in rebuttal).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SMORe instead of GoFAR | Offline GCRL · GCRL tasks | performance | Helped |
| Replace GoFAR's learned discriminator with the dataset reward (ablation) | GoFAR · Sparse-reward GCRL | performance | Hurt |
| Add HER augmentation (ablation) | SMORe · Offline GCRL tasks | performance | Helped |
| Add in-sample maximization to GoFAR (ablation) | GoFAR · Offline GCRL tasks | performance | No clear effect |
| Use only random, suboptimal data (ablation) | Offline GCRL methods · Random-data setting | success | Hurt |
| Remove techniques that constrain the learned policy to the offline data distribution (ablation) | SMORe · Offline GCRL tasks | policy performance | Hurt |
Surprising findings
- GoFAR with reward instead of discriminator has poor performance, contradicting the claim that the discriminator is optional.
- Using 0 expert data or without HER still yields results competitive with GoFAR.
Limitations
- Claimed discriminator-free status is disputed by reviewers since the S-function acts similarly to a discriminator.
- Requires two networks (S and M), more complexity than single-network contrastive RL.
Experimental setup
Datasets: Fetch benchmark offline datasets (random, medium, expert) Sawyer benchmark offline datasets locomotion goal-conditioned datasets (Ant/Cheetah TgtVel)
Benchmarks: Fetch Sawyer AntTgtVel CheetahTgtVel
Metrics: discounted return success rate
Baselines: GoFAR GCSL WGCSL contrastive RL IQL HER
Builds on: GoFAR Implicit Q-learning Advantage-weighted regression Hindsight experience replay DualDICE
Closest papers by meaning
- OGBench: Benchmarking Offline Goal-Conditioned RL (ICLR 2025 · similarity 0.92)
- Stitching Sub-trajectories with Conditional Diffusion Model for Goal-Conditioned Offline RL (AAAI 2024 · similarity 0.91)
- Flattening Hierarchies with Policy Bootstrapping (NeurIPS 2025 · similarity 0.89)
- Conservative Offline Goal-Conditioned Implicit V-Learning (ICML 2025 · similarity 0.89)
- Unifying Goal-Conditioned RL and Unsupervised Skill Learning via Control-Maximization (NeurIPS 2026 · similarity 0.88)
- The Generalization Gap in Offline Reinforcement Learning (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.