Research map / Reinforcement learning / Visual and representation RL
Goal-conditioned policy learning
Train goal-reaching policies with learned goal representations, discriminative rewards, and pretraining.
66 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in Visual and representation RL, keyword cluster “goal · conditioned · hierarchical”
State History Tried Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
17 papers in 2023–24 and 13 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 3.0% to 1.9%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2018 · Visual Reinforcement Learning with Imagined Goals (NeurIPS 2018 · 307 citations)
- 2019 · Curriculum-guided Hindsight Experience Replay (NeurIPS 2019 · 80 citations)
- 2020 · Skew-Fit: State-Covering Self-Supervised Reinforcement Learning (ICML 2020 · 58 citations)
- 2021 · Learning with AMIGo: Adversarially Motivated Intrinsic Goals (ICLR 2021 · 50 citations)
- 2022 · Goal Misgeneralization in Deep Reinforcement Learning (ICML 2022 · 21 citations)
- 2023 · VIP: Towards Universal Visual Reward and Representation via Value-Implicit Pre-Training (ICLR 2023 · 35 citations)
- 2024 · Unveiling the Significance of Toddler-Inspired Reward Transition in Goal-Oriented Reinforcement Learning (AAAI 2024 · 3 citations)
- 2025 · Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 2025 · 0 citations)
- 2026 · Adaptive Quasimetric Mapping : Principled Topological Abstraction for Robust Offline Goal-Conditioned Navigation (ICML 2026 · 0 citations)
Earliest papers here
- Visual Reinforcement Learning with Imagined Goals (NeurIPS 2018 · 307 citations)
- Universal Planning Networks: Learning Generalizable Representations for Visuomotor Control (ICML 2018 · 58 citations)
- Unsupervised Learning of Goal Spaces for Intrinsically Motivated Goal Exploration (ICLR 2018 · 57 citations)
What has been tried
53 papers here report 502 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 10%, hurt 17%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| vary the S2D transition timing across C1, C2, and C3 | LunarLander · performance | 138.71\stdv{3.71}; 63.40\stdv{160.55}; 168.88\stdv{23.66} | Mixed | Unveiling the Significance of Toddler-Inspired Reward Transition in Goal-Oriented Reinforcement Learning (AAAI 2024 · 3 citations) |
| Add dual-path global and local features | FiveK Random-Target · PSNR | Helped | Goal Conditioned Reinforcement Learning for Photo Finishing Tuning (NeurIPS 2024 · 3 citations) | |
| set the number of navigation subgoals Ns to 3 instead of 2 | evaluated tasks · training performance | Hurt | Learning World Models for Unconstrained Goal Navigation (NeurIPS 2024 · 2 citations) | |
| Pre-train the cDFA encoder on RAD cDFAs | RAD cDFAs, Letterworld · task learning | Helped | Compositional Automata Embeddings for Goal-Conditioned Reinforcement Learning (NeurIPS 2024 · 1 citation) | |
| add the final latent goal z_g as policy context | LIBERO-LONG · Avg. Success | 77.3 -> 83.3 | Helped | Efficient Robotic Policy Learning via Latent Space Backward Planning (ICML 2025 · 0 citations) |
| Add temporal-distance reward to FOUNDER instead of using cosine similarity | DMC and Kitchen overall · normalized test performance | 0.52 -> 0.81 | Helped | FOUNDER: Grounding Foundation Models in World Models for Open-Ended Embodied Decision Making (ICML 2025 · 0 citations) |
| Add reached-goal history to the skill decoder | MetaWorld MT50 Near-Optimal and Sub-Optimal · average episode return | 2952.1 (± 92.3) -> 3159.1 (± 71.0); 2041.5 (± 83.5) -> 2123.8 (± 78.1) | Helped | Goal-Oriented Skill Abstraction for Offline Multi-Task Reinforcement Learning (ICML 2025 · 0 citations) |
| remove latent space shaping | Coffee Avg. · success rate | 0.77 -> 0.31 | Hurt | $\textit{Hyper-GoalNet}$: Goal-Conditioned Manipulation Policy Learning with HyperNetworks (NeurIPS 2025 · 0 citations) |
| add temporal difference loss L_TD | Maze · Mean absolute travel-time error | 0.21 -> 0.08 | Helped | Physics-informed Temporal Difference Metric Learning for Robot Motion Planning (ICLR 2025 · 0 citations) |
| Condition the policy on the time horizon rather than omit horizon conditioning | Offline goal-conditioned control tasks · performance | Helped | Learning to Reach Goals via Diffusion (ICML 2024 · 0 citations) | |
| Use a separate PCLaSt map ψ instead of applying contrastive loss directly to ACRO representation φ | Maze-Hallway · score | 89.3 ± 8.1 -> 97.8 ± 4.9 | Helped | PcLast: Discovering Plannable Continuous Latent States (ICML 2024 · 0 citations) |
| Use a maximal prompt length of 5 rather than 1 | MazeRunner-15, MazeRunner-30, and Kitchen · average return | -2.63±0.00 -> -0.41±0.10; -27.06±0.00 -> -14.21±0.19; 0.09±0.00 -> 0.63±0.02 | Helped | Pre-Trained Multi-Goal Transformers with Prompt Optimization for Efficient Online Adaptation (NeurIPS 2024 · 0 citations) |
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 30 papers; from OpenAlex and the papers' affiliations).
- University of Pennsylvania · 3 papers
- Peking University · 3 papers
- Shanghai AI Laboratory · 3 papers
- Massachusetts Institute of Technology · 3 papers
- Meta · 2 papers
- McGill University · 2 papers
- Tsinghua University · 2 papers
- National University of Defense Technology · 2 papers
- Rutgers University · 2 papers
Use it
Download BibTeX of 40 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Graph search and compositional goals (55 papers)
- Subgoal generation for hierarchical RL (35 papers)
Elsewhere in the topic
- In-context and adaptive RL (202 papers)
- Scalable RL systems and environments (179 papers)
- Probabilistic and autonomous RL (177 papers)
- RL for combinatorial routing (158 papers)
- Learned state and action representations (148 papers)
- Dexterous manipulation transfer (141 papers)
- RL with tree search for discovery (121 papers)
- Inferring decision-making from behavior (93 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).