Research map / Reinforcement learning / Imitation learning
SeMAIL: Eliminating Distractors in Visual Imitation via Separated Models
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 |
|---|---|---|---|
| Add action-conditioned task-relevant model with action-free task-irrelevant (AcT) dynamics (ablation) | SeMAIL · Six visual control tasks (DMC with video backgrounds, Car Racing) | final episodic return: Walker Walk 900.1 vs 124.9 without AcT; Walker Run 463.4 vs 100.0 | Helped |
| Add background-only reconstruction (BoR) loss (ablation) | SeMAIL · Walker Walk / Walker Run (DMC with video backgrounds) | final episodic return: Walker Walk 900.1 vs 93.7; Walker Run 463.4 vs 337.5 | Helped |
| Add background-only reconstruction (BoR) loss (ablation) | SeMAIL · Car Racing | final episodic return: SeMAIL (No BoR) 942.9 vs SeMAIL 901.2 | Hurt |
| Add BoR loss (no-BoR variant) on non-overlap backgrounds (ablation) | SeMAIL · Walker Run, Cheetah Run, Finger Spin with non-overlap backgrounds | scaled return: Walker Run: No BoR 0.84 vs full 0.38; Cheetah Run: No BoR 0.38 vs full 0.86; Finger Spin: No BoR 0.75 vs full 0.82 | Mixed |
| Train separated models only on agent data (no expert data in model learning) (ablation) | SeMAIL · Walker Walk / Walker Run / Cheetah Run / Finger Spin / Hopper Hop | final episodic return: Walker Walk 900.1 vs 62.8; Walker Run 463.4 vs 63.7 | Hurt |
| Full SeMAIL vs V-MAIL (task-relevant latent imagination vs full latent) | model-based AIL · Walker Walk, Walker Run, Cheetah Run, Finger Spin, Hopper Hop | final episodic return: Walker Walk 900.1 vs 314.9; Cheetah Run 217.9 vs 74.6; Hopper Hop 90.5 vs 0.2 | Helped |
| Full SeMAIL vs V-MAIL on Car Racing | model-based AIL · Car Racing | final episodic return: SeMAIL 901.2 vs V-MAIL 537.8 (SeMAIL higher, but V-MAIL early returns were high) | Hurt |
| Full SeMAIL vs DA-DAC and DisentanGAIL | visual imitation with distractor backgrounds · Walker Walk, Walker Run, Cheetah Run, Finger Spin, Hopper Hop | final episodic return: e.g. Walker Walk 900.1 vs 25.4 (DA-DAC) and 61.0 (DisentanGAIL) | Helped |
| Full SeMAIL vs DisentanGAIL on non-overlap backgrounds | visual imitation with unseen backgrounds · Finger Spin (non-overlap backgrounds) | scaled return: DisentanGAIL 1.47 vs SeMAIL 0.82 | Hurt |
| Using discriminator pseudo-reward to separate task-relevant information in prior Non-IL methods (ablation) | Non-IL methods with noisy observations · Walker Run, Finger Spin, Hopper Hop | episodic return: Walker Run 50.4 vs random 29.0; Finger Spin 3.3 vs 2.1 | No clear effect |
Closest papers by meaning
- Visual Adversarial Imitation Learning using Variational Models (NeurIPS 2021 · similarity 0.90)
- Imitation Learning via Kernel Mean Embedding (AAAI 2018 · similarity 0.89)
- Diffusion Imitation from Observation (NeurIPS 2024 · similarity 0.87)
- Adversarial Imitation Learning via Boosting (ICLR 2024 · similarity 0.87)
- Diffusing States and Matching Scores: A New Framework for Imitation Learning (ICLR 2025 · similarity 0.86)
- Robust Visual Imitation Learning with Inverse Dynamics Representations (AAAI 2024 · similarity 0.86)
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.