Gymnasium
An open-source reinforcement learning environment interface library that provides standardized APIs for game and simulation environments. It enables researchers worldwide to compare algorithm performance on unified benchmarks and serves as a foundational infrastructure for reinforcement learning research.
Core Facts
Timeline (last 90 days)
The researchers conducted computational experiments on two types of benchmarks: the Feynman dataset (physical formula regression tasks) and Gymnasium reinforcement learning benchmarks.
MuJoCo has become one of the core simulation backends in the OpenAI Gym / Gymnasium ecosystem.
The environment follows the Gymnasium API standard; the observation space is a 13,977-dimensional float vector and incorporates the historical record of cards played by the opponent.
gym-pybullet-drones provides a native Gymnasium interface that is directly compatible with mainstream RL libraries such as Stable-Baselines3 and RLlib
Bot-friendly reinforcement learning environments typically follow standard interface specifications similar to OpenAI Gym/Gymnasium, providing standard methods such as observation, action, reward, and step/reset
MuJoCo is the default physics backend for classic control tasks in OpenAI Gym/Gymnasium.
Gym/Gymnasium defines a unified interaction interface reset(), step(), render() for reinforcement learning environments, but standardization stops at the environment interaction API and does not cover higher-level engineering processes such as training loop management and experiment configuration.
OpenAI launched Gym in 2016, which has since evolved into Gymnasium, now maintained by the Farama Foundation.
Gymnasium将原版Gym中的done信号明确拆分为terminated（任务自然结束）和truncated（因时间限制截断）
2022年OpenAI将Gym的维护工作移交给Farama基金会，并更名为Gymnasium
11 more timeline events
All Facts (20)
Gymnasium, maintained by the Farama Foundation, succeeded OpenAI Gym and defines the environment interface standard
70%UnverifiedThe researchers conducted computational experiments on two types of benchmarks: the Feynman dataset (physical formula regression tasks) and Gymnasium reinforcement learning benchmarks.
50%UnverifiedMuJoCo has become one of the core simulation backends in the OpenAI Gym / Gymnasium ecosystem.
50%UnverifiedThe environment follows the Gymnasium API standard; the observation space is a 13,977-dimensional float vector and incorporates the historical record of cards played by the opponent.
50%Unverifiedgym-pybullet-drones provides a native Gymnasium interface that is directly compatible with mainstream RL libraries such as Stable-Baselines3 and RLlib
50%UnverifiedBot-friendly reinforcement learning environments typically follow standard interface specifications similar to OpenAI Gym/Gymnasium, providing standard methods such as observation, action, reward, and step/reset
50%UnverifiedMuJoCo is the default physics backend for classic control tasks in OpenAI Gym/Gymnasium.
50%UnverifiedOpenAI launched Gym in 2016, which has since evolved into Gymnasium, now maintained by the Farama Foundation.
50%UnverifiedGym/Gymnasium defines a unified interaction interface reset(), step(), render() for reinforcement learning environments, but standardization stops at the environment interaction API and does not cover higher-level engineering processes such as training loop management and experiment configuration.
50%UnverifiedGymnasium将原版Gym中的done信号明确拆分为terminated（任务自然结束）和truncated（因时间限制截断）
50%Unverified2022年OpenAI将Gym的维护工作移交给Farama基金会，并更名为Gymnasium
50%UnverifiedGymnasium提供统一的环境接口（reset、step、render），覆盖CartPole、MountainCar、Atari游戏、MuJoCo物理仿真等环境
50%Unverified作者为《雨世界》环境添加了Gymnasium封装，使其可直接使用Stable-Baselines3中的算法
50%UnverifiedGym/Gymnasium 环境接口的核心设计是任何环境只需实现 reset() 和 step() 两个方法即可与算法库对接
50%Unverified在 Apple M1 Max 上，reinfors 单线程步进速度约为 Gymnasium 的 20 倍（3,850 vs 195 步/秒）
50%Unverifiedreinfors 直接输出 CHW 格式的观测数据（PyTorch卷积层期望格式），可通过 np.transpose(obs, (2, 0, 1)) 转换到标准 Gymnasium 的 HWC 格式
50%UnverifiedGymnasium的AsyncVectorEnv基于Python的multiprocessing模块实现并行，每个子环境运行在独立进程中
50%Unverified在 AMD EPYC EC2 实例上，reinfors 单线程步进速度约为 Gymnasium 的 14 倍（2,069 vs 148 步/秒）
50%UnverifiedGymnasium是OpenAI Gym的社区维护后继版本，已成为强化学习环境的事实标准接口
50%UnverifiedGymnasium的step(action)方法接收动作后返回五元组（observation, reward, terminated, truncated, info）
50%