Research map / Reinforcement learning / Imitation learning
Coherent Soft Imitation Learning
Joe Watson, Sandy Huang, Nicolas Heess
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CSIL instead of the strongest listed baseline (GAIL) | Tabular inverse optimal control · Dense Nominal | normalized performance: 0.251 -> 0.257 | Improved |
| Use CSIL instead of the strongest listed baseline (GAIL) | Tabular inverse optimal control · Dense Windy | normalized performance: 0.111 -> 0.107 | Hurt |
| Use CSIL instead of the strongest baselines | Tabular inverse optimal control · Sparse Nominal | normalized performance: 1.237 -> 1.237 | No clear effect |
| Use CSIL instead of the strongest listed baseline (Classifier) | Tabular inverse optimal control · Sparse Windy | normalized performance: 0.044 -> 0.044 | No clear effect |
| Pretrain IQLearn or PPIL with behavioral cloning before imitation learning (ablation) | IQLearn and PPIL · Online continuous control | episode return | Mixed |
| Use negative CSIL rewards rather than positive CSIL rewards (ablation) | CSIL on robomimic tasks · Online human demonstrations | success rate | Improved |
| Use the negative CSIL reward rather than a constant negative reward (ablation) | CSIL on robomimic tasks · Online human demonstrations, fewer demonstrations | success rate | Improved |
| Use a faithful heteroscedastic regression loss instead of NLLH (ablation) | Heteroscedastic regression policies · Regression toy problem | mean prediction error | Improved |
| Use a faithful heteroscedastic regression loss instead of NLLH (ablation) | Heteroscedastic regression policies · Regression toy problem | uncertainty quantification | No clear effect |
| Regularize finetuning against the cloned policy instead of using the uniform prior (ablation) | CSIL · Continuous control | policy stability | Improved |
| Refine the coherent reward using additional replay data (ablation) | CSIL · Continuous control, especially fewer demonstrations | downstream performance | Improved |
| Use an unrefined reward rather than minimax-refining the coherent reward (ablation) | CSIL · Continuous control | learning stability | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- IQLearn — IQ-Learn: Inverse soft-Q Learning for Imitation (NeurIPS 2021) · better on 1 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning (NeurIPS 2024 · similarity 0.92)
- EvIL: Evolution Strategies for Generalisable Imitation Learning (ICML 2024 · similarity 0.91)
- Interactive and Hybrid Imitation Learning: Provably Beating Behavior Cloning (NeurIPS 2025 · similarity 0.91)
- Reward-free World Models for Online Imitation Learning (ICML 2025 · similarity 0.90)
- Hybrid Inverse Reinforcement Learning (ICML 2024 · similarity 0.90)
- Learning When to Stop: Selective Imitation Learning Under Arbitrary Dynamics Shift (NeurIPS 2026 · similarity 0.90)
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.
