Research map / Reinforcement learning / Imitation learning
State Alignment-based Imitation Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 state alignment instead of action imitation under modified agent dynamics | SAIL · Modified Ant environments | reward: 1042.45 ± 75.13 -> 3305.71 ± 67.21 (DisabledAnt); 550.6 ± 77.62 -> 5608.47 ± 57.67 (HeavyAnt) | Helped |
| use state alignment under modified dynamics | SAIL · Modified Swimmer environments | reward: 287.71 ± 2.31 vs 283.42 ± 3.69 (DisabledSwimmer); 342.61 ± 6.14 vs 277.99 ± 3.41 (LightSwimmer); 286.4 ± 3.2 vs 301.27 ± | Mixed |
| use SAIL instead of BC | SAIL · Hopper-v2, 10 demos | reward: 1318.76 ± 804.36 -> 3539.56 ± 130.36 | Helped |
| use SAIL instead of BC | SAIL · Hopper-v2, 50 demos | reward: 3525.87 ± 160.74 -> 3614.19 ± 150.74 | No clear effect |
| use SAIL instead of BC | SAIL · HalfCheetah-v2, 50 demos | reward: 4813.20 ± 1949.26 -> 8817.32 ± 860.55 | Helped |
| set β to 0.2 rather than β=0.1 (ablation) | SAIL · HalfCheetah-50 | reward: 7102.29 -> 2007.86 | Hurt |
| use β=0.1 rather than β=0.2 (ablation) | SAIL · HeavyAnt-Final | reward: 282.13 -> 5208.45 | Helped |
| replace the VAE with an ordinary MLP (ablation) | SAIL · HalfCheetah-50 | reward: 7102.29 -> 4759.69 | Hurt |
| predict next actions with a VAE rather than predict next states (ablation) | SAIL · HalfCheetah-50 | reward: 4813.20 ± 1949.26 vs 230.52 ± 13.26 (β=0.1) | Hurt |
| predict next actions with a VAE rather than predict next states (ablation) | SAIL · Hopper-50 | reward: 3525.87 ± 6.74 vs 203.87 ± 14.39 (β=0.1) | Hurt |
| combine Wasserstein state-distribution reward with KL policy-prior regularization (ablation) | SAIL · HalfCheetah-v2 and Humanoid-v2, 20 demos | reward | Helped |
| use the Wasserstein objective without the KL prior (ablation) | SAIL · HalfCheetah-v2 and Humanoid-v2, 20 demos | reward | Hurt |
Closest papers by meaning
- State-only Imitation with Transition Dynamics Mismatch (ICLR 2020 · similarity 0.91)
- Rethinking Inverse Reinforcement Learning: from Data Alignment to Task Alignment (NeurIPS 2024 · similarity 0.90)
- Robust Visual Imitation Learning with Inverse Dynamics Representations (AAAI 2024 · similarity 0.90)
- Interactive and Hybrid Imitation Learning: Provably Beating Behavior Cloning (NeurIPS 2025 · similarity 0.89)
- Agnostic Interactive Imitation Learning: New Theory and Practical Algorithms (ICML 2024 · similarity 0.89)
- Reward-free World Models for Online Imitation Learning (ICML 2025 · similarity 0.89)
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.