Research map / Reinforcement learning / Visual and representation RL
EnvPool: A Highly Parallel Reinforcement Learning Environment Execution Engine
Jiayi Weng, Min Lin, Shengyi Huang, Bo Liu, Denys Makoviichuk, Viktor Makoviychuk, Zichen Liu, Yufan Song, Ting Luo, Yukun Jiang, Zhongwen Xu, Shuicheng YAN
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 |
|---|---|---|---|
| Replace Python subprocess environment execution with EnvPool (async) | RL environment simulation · Laptop, Atari | FPS: 15,863 -> 49,439 | Improved |
| Replace Python subprocess environment execution with EnvPool (async) | RL environment simulation · Laptop, MuJoCo | FPS: 36,586 -> 105,126 | Improved |
| Replace Sample Factory with EnvPool (async) | RL environment simulation · Atari and MuJoCo across Laptop and Workstation | FPS: 28,216 -> 49,439; 62,510 -> 105,126; 138,847 -> 200,428; 309,264 -> 582,446 | Improved |
| Use NUMA + async execution rather than Sample Factory | RL environment simulation · Atari and MuJoCo, DGX-A100 | FPS: 707,494 -> 1,069,922; 1,573,262 -> 3,134,287 | Improved |
| Use asynchronous rather than synchronous execution (ablation) | EnvPool · Atari and MuJoCo across Laptop, Workstation, and DGX-A100 | FPS: 37,396 -> 49,439; 66,622 -> 105,126; 133,824 -> 200,428; 380,950 -> 582,446; 427,851 -> 891,286; 949,787 -> 2,363,864 | Improved |
| Replace Python with EnvPool for a single environment | Environment simulation · Atari Pong-v5, Laptop | FPS: 4,891 -> 7,887 | Improved |
| Replace Python with EnvPool for a single environment | Environment simulation · MuJoCo Ant-v3, Laptop | FPS: 12,325 -> 15,641 | Improved |
| Replace Python with EnvPool for a single environment | Environment simulation · DeepMind Control cheetah run, Laptop | FPS: 6,235 -> 11,636 | Improved |
| Use EnvPool (Sync) instead of For-loop execution | CleanRL PPO on Atari Breakout · Atari Breakout, N=8 | end-to-end training time: ~200 minutes -> approximately 73 minutes | Improved |
| Use EnvPool instead of DummyVecEnv | Acme PPO · MuJoCo HalfCheetah-v3 | wall time to similar final episode return | Improved |
| Use EnvPool instead of Ray parallel environments | rl_games PPO · Atari and MuJoCo example runs | training speed | Improved |
| Tune the number of parallel environments while keeping environment-interaction budget comparable | Acme PPO with EnvPool · MuJoCo HalfCheetah-v3 | training time | Improved |
Closest papers by meaning
- NAVIX: Scaling MiniGrid Environments with JAX (NeurIPS 2025 · similarity 0.87)
- SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores (ICLR 2024 · similarity 0.87)
- Jumanji: a Diverse Suite of Scalable Reinforcement Learning Environments in JAX (ICLR 2024 · similarity 0.85)
- Craftax: A Lightning-Fast Benchmark for Open-Ended Reinforcement Learning (ICML 2024 · similarity 0.85)
- Efficient Reinforcement Learning Through Adaptively Pretrained Visual Encoder (AAAI 2025 · similarity 0.85)
- JaxMARL: Multi-Agent RL Environments and Algorithms in JAX (NeurIPS 2024 · 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.
