Research map / Reinforcement learning / Imitation learning
Hierarchical Few-Shot Imitation with Skill Transition 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 |
|---|---|---|---|
| Use FIST rather than the strongest reported baseline | PointMaze · Left, Right, and Bottom blocked regions | episode length: 363.87 $\pm$ 18.73 vs 1089.76 $\pm$ 173.74; 571.21 $\pm$ 38.82 vs 1918.99 $\pm$ 43.65; 359.82 $\pm$ 3.62 vs 1127.47 $\pm | Helped |
| Use FIST rather than the strongest reported baseline | PointMaze · Left, Right, and Bottom blocked regions | success rate: 0.99 $\pm$ 0.03 vs 0.74 $\pm$ 0.11; 0.91 $\pm$ 0.07 vs 0.07 $\pm$ 0.06; 1.0 $\pm$ 0.0 vs 0.87 $\pm$ 0.10 | Helped |
| Use FIST rather than the strongest reported baseline | AntMaze · Left, Right, and Bottom blocked regions | success rate: 0.32 $\pm$ 0.04 vs 0.0 $\pm$ 0.0; 0.22 $\pm$ 0.12 vs 0.0 $\pm$ 0.0; 0.21 $\pm$ 0.07 vs 0.12 $\pm$ 0.07 | Helped |
| Use FIST rather than SPiRL | KitchenRobot · Top Burner, Kettle, and Slide Cabinet excluded tasks | episode reward: 3.6 \pm 0.16 vs 2.1 \pm 0.48; 3.5 \pm 0.3 vs 1.9 \pm 0.09; 4.0 \pm 0.0 vs 3.3 \pm 0.38 | Helped |
| Use FIST rather than SPiRL | KitchenRobot · Microwave excluded task | episode reward: 2.3 \pm 0.5 vs 2.3 \pm 0.5 | No clear effect |
| Add learned contrastive distance instead of Euclidean raw-state distance (ablation) | FIST · KitchenRobot tasks with four excluded-object settings | episode reward: 3.3 $\pm$ 0.15 vs 3.6 \pm 0.16; 2.8 $\pm$ 0.42 vs 2.3 \pm 0.5; 3.0 $\pm$ 0.32 vs 3.5 \pm 0.3; 1.3 $\pm$ 0.14 vs 4.0 \pm | Helped |
| Add future-conditioned inverse skill prior rather than current-state prior (ablation) | SPiRL with semi-parametric demonstration lookup · KitchenRobot | episode reward: SPiRL (closest): 0.0 $\pm$ 0.0, 2.8 $\pm$ 0.44, 0.0 $\pm$ 0.0, 2.4 $\pm$ 0.51; FIST: 3.6 \pm 0.16, 2.3 \pm 0.5, 3.5 \pm | Helped |
| Add downstream fine-tuning (ablation) | FIST · KitchenRobot OOD tasks | episode reward: FIST: 3.6 \pm 0.16, 2.3 \pm 0.5, 3.5 \pm 0.3, 4.0 \pm 0.0; FIST-no-FT: 2.0 $\pm$ 0.0, 0.0 $\pm$ 0.0, 1.0 $\pm$ 0.0, 2.0 | Helped |
| Add offline skill pre-training (ablation) | FIST · KitchenRobot OOD tasks | episode reward: FIST: 3.6 \pm 0.16, 2.3 \pm 0.5, 3.5 \pm 0.3, 4.0 \pm 0.0; FIST-no-pretrain: 0.5 $\pm$ 0.16, 0.7 $\pm$ 0.15, 0.0 $\pm$ 0 | Helped |
| Use inverse skill dynamics rather than atomic-action Goal-BC (ablation) | FIST · KitchenRobot OOD tasks | episode reward: 3.6 \pm 0.16 vs 0.0 $\pm$ 0.0; 2.3 \pm 0.5 vs 1.2 $\pm$ 0.3; 3.5 \pm 0.3 vs 1.8 $\pm$ 0.44; 4.0 \pm 0.0 vs 0.9 $\pm$ 0.1 | Helped |
| Use an oracle future-state lookup instead of learned contrastive lookup (ablation) | FIST · PointMaze Left, Right, and Bottom | episode length: 363.87 $\pm$ 18.73 vs 236.00 $\pm$ 1.02; 571.21 $\pm$ 38.82 vs 280.93 $\pm$ 5.61; 359.82 $\pm$ 3.62 vs 269.89 $\pm$ 3.75 | Helped |
| Resample the skill every 10 steps instead of every step (ablation) | FIST · PointMaze | episode length: Left: 329.09 $\pm$ 5.62 vs 338.04 $\pm$ 9.80; Right: 440.47 $\pm$ 50.81 vs 465.05 $\pm$ 52.38; Bottom: 491.67 $\pm$ 78.4 | Mixed |
Closest papers by meaning
- OSVI-WM: One-Shot Visual Imitation for Unseen Tasks using World-Model-Guided Trajectory Generation (NeurIPS 2025 · similarity 0.91)
- GO-DICE: Goal-Conditioned Option-Aware Offline Imitation Learning via Stationary Distribution Correction Estimation (AAAI 2024 · similarity 0.90)
- Zero-Shot Offline Imitation Learning via Optimal Transport (ICML 2025 · similarity 0.89)
- Robust Policy Learning via Offline Skill Diffusion (AAAI 2024 · similarity 0.89)
- Meta-Controller: Few-Shot Imitation of Unseen Embodiments and Tasks in Continuous Control (NeurIPS 2024 · similarity 0.89)
- LLM-based Skill Diffusion for Zero-shot Policy Adaptation (NeurIPS 2024 · 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.