Research map / Reinforcement learning / Imitation learning
Difference-Aware Retrieval Polices for Imitation Learning
Quinn Pfeifer, Ethan Pronovost, Paarth Shah, Khimya Khetarpal, Siddhartha Srinivasa, Abhishek Gupta
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 difference-aware retrieval policy instead of plain behavior cloning | BC (MLP) · Robosuite and RoboCasa (low-dim state) | success rate (%): Stack 47 -> 72 | Improved |
| use DARP instead of BC on vision-based tasks (R3M features) | BC · Robosuite Peg | success rate (%): 17 -> 52 | Improved |
| add neighbor difference vectors (query vs neighbor state) to the predictor input (ablation) | DARP · Robosuite ablation | success rate | Improved |
| use permutation-invariant aggregation of neighbor predictions (ablation) | DARP · Robosuite ablation | success rate | Improved |
| include neighbor actions in predictor input (ablation) | DARP · Robosuite ablation | success rate | Improved |
| select neighbors by distance rather than at random (ablation) | DARP · Robosuite ablation | success rate | Improved |
| implicit neighbor aggregation in the architecture (DARP) instead of explicit smoothness loss (MRIL) (ablation) | MRIL · MuJoCo HalfCheetah | return: 701.58 -> 5515.41 | Improved |
| use pure non-parametric retrieval (R&P, LWR) | BC · MuJoCo | return: Hopper 2313.65 -> 711.82 (R&P) | Hurt |
| replace Euclidean distance with cosine similarity for neighbor retrieval (ablation) | DARP · Robosuite Stack (R3M) | success rate (%): 75 -> 70 | No clear effect |
| combine DARP with a diffusion policy backbone (ablation) | DARP (MLP) · Robosuite Stack | success rate (%): 62 -> 82 | Improved |
| use DARP with GMM head on multimodal Push-T | BC · Push-T | performance: 22% improvement over BC | Improved |
| use DARP on discontinuous modified D4RL Umaze | BC · D4RL Umaze (modified) | success rate: 25% -> 57% | Improved |
Closest papers by meaning
- Videos are Sample-Efficient Supervisions: Behavior Cloning from Videos via Latent Representations (NeurIPS 2025 · similarity 0.90)
- Interactive and Hybrid Imitation Learning: Provably Beating Behavior Cloning (NeurIPS 2025 · similarity 0.90)
- Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning (NeurIPS 2024 · similarity 0.89)
- Rethinking Latent Redundancy in Behavior Cloning: An Information Bottleneck Approach for Robot Manipulation (ICML 2025 · similarity 0.89)
- How to Leverage Diverse Demonstrations in Offline Imitation Learning (ICML 2024 · similarity 0.89)
- Split the Differences, Pool the Rest: Provably Efficient Multi-Objective Imitation (NeurIPS 2026 · 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.
