Research map / Reinforcement learning / Imitation learning
Jump-Start Reinforcement Learning
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 JSRL curriculum of guide-steps on top of IQL (cold-start) | IQL · D4RL Ant Maze and Adroit, offline 1k-100k | normalized score | Helped |
| add JSRL curriculum to IQL on Ant Maze with 1m offline transitions | IQL · D4RL Ant Maze 1m offline | normalized score | No clear effect |
| add JSRL curriculum to IQL on Adroit pen/door/relocate with 100k offline | IQL · Adroit 100k offline | normalized score: door 50.2 IQL vs 28.5 JSRL curriculum | Hurt |
| add JSRL curriculum to QT-Opt with BC guide on Instance and Indiscriminate Grasping | QT-Opt · Robotic grasping 20 demos | success rate: Indiscriminate 0.00 to 0.91; Instance 0.29 to 0.54 | Helped |
| add JSRL to QT-Opt in high-data grasping regime (20k demos) | QT-Opt · Indiscriminate Grasping 20k demos | success rate: 0.95 vs AW-Opt 0.93, QT-Opt 0.93 | No clear effect |
| random switching instead of curriculum for guide-steps (IQL cold-start) (ablation) | IQL+JSRL · Ant Maze 1k offline | normalized score: umaze 10.35 random vs 15.60 curriculum; umaze-diverse 1.90 vs 3.05 | Mixed |
| warm-start vs cold-start initialization of exploration policy (IQL+JSRL curriculum, Adroit 10k) (ablation) | IQL+JSRL · Adroit 10k offline | normalized score: door 4.33 cold vs 5.08 warm; relocate 0.55 cold vs 4.39 warm | Mixed |
| warm-start vs cold-start (IQL+JSRL random, Adroit 1k) (ablation) | IQL+JSRL · Ant Maze 1k offline | normalized score: antmaze-umaze warm 0.10 vs cold 10.35 | Hurt |
| larger performance-threshold tolerance for moving curriculum stages (5%-15%) (ablation) | JSRL curriculum on indiscriminate grasping · Indiscriminate Grasping | success rate | Hurt |
| use indiscriminate-grasping policy as guide instead of instance-grasping guide (ablation) | QT-Opt+JSRL · Instance Grasping guide generalization | success rate: slightly worse than instance guide | Hurt |
| use pre-trained simpler antmaze-play guide for diverse antmaze (IQL+JSRL) | IQL+JSRL · Ant Maze diverse generalization | normalized score | Helped |
Closest papers by meaning
- Heuristic-Guided Reinforcement Learning (NeurIPS 2021 · similarity 0.91)
- Ready Policy One: World Building Through Active Learning (ICML 2020 · similarity 0.90)
- Dual RL: Unification and New Methods for Reinforcement and Imitation Learning (ICLR 2024 · similarity 0.90)
- Enhancing Online Reinforcement Learning with Meta-Learned Objective from Offline Data (AAAI 2025 · similarity 0.90)
- Hybrid Inverse Reinforcement Learning (ICML 2024 · similarity 0.90)
- Overcoming the Sim-to-Real Gap: Leveraging Simulation to Learn to Explore for Real-World RL (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.