Research map / Reinforcement learning / Imitation learning
Learning Parameterized Skills from Demonstrations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DEPS, an end-to-end hierarchical method that learns parameterized skills (discrete skill plus continuous parameters) from multitask demonstrations using temporal variational inference, information bottlenecks, and 1D state compression, evaluated for few-shot generalization on LIBERO and MetaWorld.
Problem
Learning reusable skills from demonstrations that generalize to unseen tasks, avoiding degenerate solutions where a high-capacity low-level policy memorizes task-specific behavior and skill abstractions are not learned.
Approach
Jointly train a discrete skill selector, a per-skill continuous parameter policy, and a low-level subpolicy, where the subpolicy sees only a learned 1D projected robot state conditioned on the skill parameters, forcing skills to carry task-relevant information and act as trajectory indices.
Key results
- On LIBERO-OOD, DEPS reaches 0.34 mean success vs 0.15 for BC and 0.10 for PRISE.
- On LIBERO-3-shot, DEPS achieves 0.26 mean success vs 0.11 for BC and 0.07 for PRISE.
- Discrete-only and continuous-only ablations reach 0.06 and 0.15 mean success on LIBERO-OOD, well below the full method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace 1D state compression with 2D state compression (ablation) | DEPS · LIBERO-OOD | Mean Success Rate: 0.34 ± 0.08 -> 0.13 ± 0.12 | Hurt |
| replace 1D state compression with 3D state compression (ablation) | DEPS · LIBERO-OOD | Mean Success Rate: 0.34 ± 0.08 -> 0.05 ± 0.04 | Hurt |
| remove state compression (ablation) | DEPS · LIBERO-OOD | Mean Success Rate: 0.34 ± 0.08 -> 0.19 ± 0.05 | Hurt |
| increase maximum discrete skills from 10 to 20 (ablation) | DEPS · LIBERO-OOD | Mean Success Rate: 0.34 ± 0.08 -> 0.47 ± 0.01 | Helped |
| remove continuous skill parameters (ablation) | DEPS · LIBERO-OOD | Mean Success: 0.34 ± 0.08 -> 0.06 ± 0.02 | Hurt |
| remove discrete skills (ablation) | DEPS · LIBERO-OOD | Mean Success: 0.34 ± 0.08 -> 0.15 ± 0.16 | Hurt |
| use parameterized skills with 1D state compression instead of BC | LIBERO-OOD · LIBERO-OOD | Mean Success: 0.15 -> 0.34 | Helped |
| pretrain with DEPS instead of BC-Untrained | LIBERO-OOD · LIBERO-OOD | Mean Success: 0.01 | Helped |
| use parameterized skills with 1D state compression instead of BC | LIBERO-3-shot · LIBERO-3-shot | Mean Success: 0.11 -> 0.26 | Helped |
| use parameterized skills with 1D state compression instead of PRISE | LIBERO-3-shot · LIBERO-3-shot | Mean Success: 0.07 -> 0.26 | Helped |
| use DEPS rather than BC | MW-Vanilla (20 pretraining epochs) · MW-Vanilla | Mean Success: 0.35 ± 0.02 -> 0.40 ± 0.02 | Helped |
| use DEPS rather than BC | MW-PRISE (20 pretraining epochs) · MW-PRISE | Mean Success: 0.24 ± 0.01 -> 0.30 ± 0.04 | Helped |
Surprising findings
- Removing the skill index from the subpolicy input improved performance, and aggressive compression of state to 1D outperformed 2D, 3D and uncompressed variants.
- Performance on the in-distribution LIBERO-10 benchmark is weaker than on held-out sets, which the authors attribute to the compression trade-off.
Limitations
- Aggressive 1D state compression may sacrifice reconstruction of long-horizon in-distribution trajectories (weaker on LIBERO-10).
- Restriction of one continuous parameter set per skill per trajectory may not suit longer-horizon tasks.
- Results are reported with few seeds (3-5) and high variance on some settings.
Experimental setup
Datasets: LIBERO-90 LIBERO-10 MetaWorld-v2
Benchmarks: LIBERO-OOD LIBERO-10 LIBERO-3-shot MW-Vanilla MW-PRISE
Metrics: mean success mean highest success
Baselines: BC BC-Untrained PRISE
Builds on: temporal variational inference PRISE LIBERO MetaWorld
Closest papers by meaning
- Language-guided Skill Learning with Temporal Variational Inference (ICML 2024 · similarity 0.89)
- Latent Diffusion Planning for Imitation Learning (ICML 2025 · similarity 0.85)
- SkillDiffuser: Interpretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Execution (CVPR 2024 · similarity 0.85)
- QueST: Self-Supervised Skill Abstractions for Learning Continuous Control (NeurIPS 2024 · similarity 0.85)
- STAR: Learning Diverse Robot Skill Abstractions through Rotation-Augmented Vector Quantization (ICML 2025 · similarity 0.85)
- Goal-Oriented Skill Abstraction for Offline Multi-Task Reinforcement Learning (ICML 2025 · similarity 0.85)
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.