Research map / Reinforcement learning / Imitation learning
Demonstration-Conditioned Reinforcement Learning for Few-Shot Imitation
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 DCRL (RL-trained demonstration-conditioned policy) instead of behaviour cloning DCBC+MT | few-shot imitation policy · Meta-World ML45 hold-out tasks | return: 296.34 vs 109.25 (1 demo) | Helped |
| Use DCRL instead of DCBC+MT | few-shot imitation policy · Navigation 10 test mazes | success rate: 77% vs 68% (1 demo); 85% vs 68% (5 demos) | Helped |
| Increase number of input demonstrations from 1 to 5 | DCRL · Meta-World ML45 hold-out tasks | success rate: 48% vs 48% | No clear effect |
| Increase number of input demonstrations from 1 to 5 | DCRL · Navigation 10 test mazes | success rate: 77% to 85% | Helped |
| Cross-demonstration attention (vs averaging per-demonstration action probabilities) (ablation) | DCRL · Navigation 10 test mazes, 5 demos | success rate: consistent advantage on all 10 test mazes (per-maze figure) | Helped |
| Train DCRL with demonstrations from a different robot (AMBIDEX) rather than Sawyer (ablation) | DCRL · Meta-World 43 tasks | success rate: 45% vs 51% (1 demo); 48% vs 51% (5 demos) | No clear effect |
| Add action noise to demonstrator (suboptimal demonstrations); DCRL trained from noisy demos outperforms expert when sigma > 2 (ablation) | DCRL · Meta-World, perturbed experts | success rate | Mixed |
| Fine-tune DCRL with BC on new task demonstrations vs DCBC+REPTILE/MTRL initialisations | DCRL fine-tuning · Meta-World ML45 fine-tuning | success rate: DCRL attains 90% success after fine-tuning on four demos | Helped |
Closest papers by meaning
- Learning from Demonstrations via Capability-Aware Goal Sampling (NeurIPS 2025 · similarity 0.90)
- Expert Proximity as Surrogate Rewards for Single Demonstration Imitation Learning (ICML 2024 · similarity 0.90)
- Imitation Learning from Observation with Automatic Discount Scheduling (ICLR 2024 · similarity 0.90)
- Limited Preference Aided Imitation Learning from Imperfect Demonstrations (ICML 2024 · similarity 0.90)
- Zero-Shot Offline Imitation Learning via Optimal Transport (ICML 2025 · similarity 0.89)
- GO-DICE: Goal-Conditioned Option-Aware Offline Imitation Learning via Stationary Distribution Correction Estimation (AAAI 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.