Research map / Reinforcement learning / Visual and representation RL
Dexterous manipulation transfer
Transfer manipulation skills with residual learning, benchmarks, and cross-embodiment data.
141 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in Visual and representation RL, keyword cluster “robot · manipulation · simulation”
State History Tried Checklist 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.
22 papers in 2023–24 and 100 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 3.8% to 14.4%.
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 · From Virtual Demonstration to Real-World Manipulation Using LSTM and MDN (AAAI 2018 · 38 citations)
- 2019 · Learning Data Manipulation for Augmentation and Weighting (NeurIPS 2019 · 66 citations)
- 2020 · Residual Force Control for Agile Human Behavior Imitation and Extended Motion Synthesis (NeurIPS 2020 · 38 citations)
- 2021 · Isaac Gym: High Performance GPU Based Physics Simulation For Robot Learning (NeurIPS 2021 · 59 citations)
- 2022 · Towards Human-Level Bimanual Dexterous Manipulation with Reinforcement Learning (NeurIPS 2022 · 28 citations)
- 2023 · Galactic: Scaling End-to-End Reinforcement Learning for Rearrangement at 100k Steps-per-Second (CVPR 2023 · 4 citations)
- 2024 · Hierarchical Diffusion Policy for Kinematics-Aware Multi-Task Robotic Manipulation (CVPR 2024 · 54 citations)
- 2025 · FlowPolicy: Enabling Fast and Robust 3D Flow-Based Policy via Consistency Flow Matching for Robot Manipulation (AAAI 2025 · 20 citations)
- 2026 · Dexterous Manipulation Transfer via Progressive Kinematic-Dynamic Alignment (AAAI 2026 · 2 citations)
Earliest papers here
- From Virtual Demonstration to Real-World Manipulation Using LSTM and MDN (AAAI 2018 · 38 citations)
- Learning Data Manipulation for Augmentation and Weighting (NeurIPS 2019 · 66 citations)
- Mirroring without Overimitation: Learning Functionally Equivalent Manipulation Actions (AAAI 2019 · 22 citations)
What has been tried
67 papers here report 632 measured changes (a component added, removed or swapped, compared with a baseline): helped 69%, no clear effect 6%, hurt 22%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| replace RK-Diffuser with RRT | RLBench overall · success rate | 94.55 -> 26.82 | Hurt | Hierarchical Diffusion Policy for Kinematics-Aware Multi-Task Robotic Manipulation (CVPR 2024 · 54 citations) |
| Increase the number of expert demonstrations | Adroit and Metaworld representative tasks · success rate | Helped | FlowPolicy: Enabling Fast and Robust 3D Flow-Based Policy via Consistency Flow Matching for Robot Manipulation (AAAI 2025 · 20 citations) | |
| Add interaction early termination | BEHAVE yogamat · Duration (s) | 36.1 -> 42.6 | Helped | InterMimic: Towards Universal Whole-Body Control for Physics-Based Human-Object Interactions (CVPR 2025 · 14 citations) |
| Add contact force to the reward | Contact-rich tasks · success rate | Helped | ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning (CVPR 2025 · 13 citations) | |
| add bi-directional regularization | chair_ingolf · task completion progress | 0.91± 0.01 -> 0.95± 0.03 | Helped | SCaR: Refining Skill Chaining for Long-Horizon Robotic Manipulation via Dual Regularization (NeurIPS 2024 · 7 citations) |
| Add reward multiplication to combine the imitation sub-rewards | GRAB · accuracy | 27.0% -> 95.4% | Helped | SkillMimic: Learning Basketball Interaction Skills from Demonstrations (CVPR 2025 · 5 citations) |
| Add force feedback inputs and force-action prediction to the residual policy | Book flipping and belt assembly · success rate | 45% higher success rate on the book task; 53% higher on the belt task | Helped | Compliant Residual DAgger: Improving Real-World Contact-Rich Manipulation with Human Corrections (NeurIPS 2025 · 5 citations) |
| Add consecutive latent-code regularization | AMASS-Test VR controller tracking · success rate | 60.8% -> 93.4% | Helped | Universal Humanoid Motion Representations for Physics-Based Control (ICLR 2024 · 4 citations) |
| Use multi-view rather than single-view image inputs | Multi-object manipulation · Sample efficiency | Helped | Entity-Centric Reinforcement Learning for Object Manipulation from Pixels (ICLR 2024 · 2 citations) | |
| Use coupled action and subskill prediction instead of decoupled prediction | Peg Insertion (0-shot) and Push Chair (unseen), state observations · success rate (%) | 47.0 -> 59.3; 36.0 -> 41.0 | Helped | Chain-of-Thought Predictive Control (ICML 2024 · 2 citations) |
| Replace adaptive tracking with one fixed tracking-factor configuration | Different motion types · Tracking error | Mixed | KungfuBot: Physics-Based Humanoid Whole-Body Control for Learning Highly-Dynamic Skills (NeurIPS 2025 · 1 citation) | |
| use hierarchical frequency-domain DCT action modeling instead of omitting DCT | Adroit, DexArt, Meta-World (10 tasks) · average success rate (%) | 61.4 -> 71.5 | Helped | FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens (NeurIPS 2025 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 44 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- GRAB · 4 of 44 papers
- Adroit · 2 of 44 papers
- Metaworld · 2 of 44 papers
- LIBERO · 2 of 44 papers
Baselines
- Diffusion Policy · 10 of 44 papers
- ACT · 6 of 44 papers
- DP3 · 5 of 44 papers
- PPO · 4 of 44 papers
- SAC · 3 of 44 papers
- pi0.5 · 3 of 44 papers
- PerAct · 2 of 44 papers
- BET · 2 of 44 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 122 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 7 papers
- Meta · 5 papers
- Shanghai Jiao Tong University · 5 papers
- Peking University · 5 papers
- Carnegie Mellon University · 4 papers
- Massachusetts Institute of Technology · 4 papers
- University of California, San Diego · 4 papers
- NVIDIA · 4 papers
- South China University of Technology · 3 papers
- Hong Kong University of Science and Technology · 3 papers
Use it
Download BibTeX of 34 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
- Skill learning and diffusion policies (86 papers)
- Vision-based robot RL (44 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)
- RL with tree search for discovery (121 papers)
- Inferring decision-making from behavior (93 papers)
- Robust visual encoders for RL (86 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).