Research map / Reinforcement learning / Imitation learning
Behavior cloning theory and models
Analyze behavior cloning and build flexible imitative or foundation-model-based policies.
68 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Imitation learning, keyword cluster “il · adversarial imitation · cloning”
State History Tried Baselines 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.
23 papers in 2023–24 and 17 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 19.3% to 14.7%.
Open questions the data points to
- Papers report Expert on Ant (return) anywhere from 300 to 4104: comparisons against it depend on whose number you take.
- Papers report Expert on HalfCheetah (return) anywhere from 560 to 4391: comparisons against it depend on whose number you take.
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
- 2016 · Model-Free Imitation Learning with Policy Optimization (ICML 2016 · 41 citations)
- 2017 · State Aware Imitation Learning (NeurIPS 2017 · 20 citations)
- 2018 · Zero-Shot Visual Imitation (ICLR 2018 · 196 citations)
- 2020 · Language-Conditioned Imitation Learning for Robot Manipulation Tasks (NeurIPS 2020 · 64 citations)
- 2021 · Learning by Watching (CVPR 2021 · 25 citations)
- 2022 · Chain of Thought Imitation with Procedure Cloning (NeurIPS 2022 · 4 citations)
- 2023 · PIRLNav: Pretraining With Imitation and RL Finetuning for ObjectNav (CVPR 2023 · 66 citations)
- 2024 · Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning (NeurIPS 2024 · 10 citations)
- 2025 · Blindfolded Experts Generalize Better: Insights from Robotic Manipulation and Videogames (NeurIPS 2025 · 0 citations)
- 2026 · On the Sample Efficiency of Inverse Dynamics Models for Semi-Supervised Imitation Learning (ICML 2026 · 0 citations)
Earliest papers here
- Model-Free Imitation Learning with Policy Optimization (ICML 2016 · 41 citations)
- State Aware Imitation Learning (NeurIPS 2017 · 20 citations)
- I See What You See: Inferring Sensor and Policy Models of Human Real-World Motor Behavior (AAAI 2017 · 7 citations)
What has been tried
56 papers here report 545 measured changes (a component added, removed or swapped, compared with a baseline): helped 55%, no clear effect 11%, hurt 25%, mixed 9%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Use a 10% expert's rather than a 110% expert's Q reference | Walker2d, expert ~110% · normalized return | 108.99±5.28 -> 78.9±49.97 | Hurt | RLIF: Interactive Imitation Learning as Reinforcement Learning (ICLR 2024 · 7 citations) |
| combine BC loss with diffusion-model loss | FetchPick · success rate | 91.6% ± 5.8% -> 97.5% ± 1.9% | Helped | Diffusion Model-Augmented Behavioral Cloning (ICML 2024 · 3 citations) |
| Remove inverse-dynamics representation learning | Hammer and Drawer Close with random masking · convergent performance | Hurt | Robust Visual Imitation Learning with Inverse Dynamics Representations (AAAI 2024 · 2 citations) | |
| Add a structure encoder | DMC average · normalized score | 31.6 -> 60.3 | Helped | Meta-Controller: Few-Shot Imitation of Unseen Embodiments and Tasks in Continuous Control (NeurIPS 2024 · 2 citations) |
| Add forward dynamics prediction | Franka Kitchen, Block Pushing, Push-T, and LIBERO Goal · relative downstream performance | 34%, 8%, 44%, 33% without forward dynamics | Helped | DynaMo: In-Domain Dynamics Pretraining for Visuo-Motor Control (NeurIPS 2024 · 2 citations) |
| Use tuned rather than fixed MCNN hyperparameters | pen-human-v1 · return | 3285 { ± 209} -> 3405 { ± 328} | Helped | Memory-Consistent Neural Networks for Imitation Learning (ICLR 2024 · 1 citation) |
| Replace the full method with DIFO-NA, using a pretrained conditional diffusion model and denoising reward without adversarial online-data training | Most evaluated tasks · learning performance | Hurt | Diffusion Imitation from Observation (NeurIPS 2024 · 1 citation) | |
| add random observation noise with P=100 | plus and ellipse test shapes · success rate | 20% -> 96%; 17% -> 79% | Helped | Blindfolded Experts Generalize Better: Insights from Robotic Manipulation and Videogames (NeurIPS 2025 · 0 citations) |
| add a warm-start phase using base-policy rollouts to train the WM, RM, and critic | Visual manipulation tasks · success rate | ≈ 20% | Helped | A Smooth Sea Never Made a Skilled SAILOR: Robust Imitation via Learning to Search (NeurIPS 2025 · 0 citations) |
| Warm-start with 3200 offline state-action pairs rather than 200 | Synthetic MDP · return | 200; 3200 offline (state, expert action) pairs | Helped | Interactive and Hybrid Imitation Learning: Provably Beating Behavior Cloning (NeurIPS 2025 · 0 citations) |
| remove exploration bonus b | Atari games · average return | 8,521 ± 1,234; 78,513 ± 11,553; 18,179 ± 21,181; 26,900 ± 52,146; 16,715 ± 23,153; 478 ± 239 | Mixed | Beyond-Expert Performance with Limited Demonstrations: Efficient Imitation Learning with Double Exploration (ICLR 2025 · 0 citations) |
| remove vector quantization | Procgen online decoding · online performance | Hurt | Learning to Act without Actions (ICLR 2024 · 0 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| Expert | Ant | return | 300 – 4104 | 2 papers, e.g. NeurIPS 2022 and AAAI 2023 |
| Expert | HalfCheetah | return | 560 – 4391 | 2 papers, e.g. NeurIPS 2022 and AAAI 2023 |
What reviewers will expect
Datasets and baselines named in the 13 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Robomimic · 2 of 13 papers
Baselines
- BC · 3 of 13 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 40 papers; from OpenAlex and the papers' affiliations).
- Nanjing University · 3 papers
- University of California, Berkeley · 3 papers
- Meta · 2 papers
- Zhejiang University · 2 papers
- Karlsruher Institut für Technologie · 2 papers
- Karlsruhe Institute of Technology · 2 papers
- Google · 2 papers
- Carnegie Mellon University · 2 papers
- National Taiwan University · 2 papers
- Korea Advanced Institute of Science and Technology · 2 papers
Use it
Download BibTeX of 37 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
- Offline imitation from imperfect data (108 papers)
- Imitation combined with RL (108 papers)
- Adversarial imitation learning (33 papers)
Elsewhere in the topic
- Reward learning from demonstrations (70 papers)
- Theory of inverse RL (44 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).