| Definition | Gymnasium-Robotics provides a collection of robotic reinforcement learning environments using the MuJoCo physics engine, designed for advanced manipulation and control tasks. |
| No of environments | 231 |
The Gymnasium-Robotics library offers a vast portfolio of robotic reinforcement learning environments built with the Gymnasium API, leveraging the powerful MuJoCo physics engine for high-fidelity simulations. The environments range from manipulation tasks using a 7-DoF Fetch robot, like reaching, pushing, and pick-and-place, to more advanced setups with the Shadow Dexterous Hand, which includes 24 degrees of freedom and variations incorporating touch sensors. These environments simulate real-world physics and control challenges, making them ideal for developing and testing robotic reinforcement learning algorithms.
In addition to the single-agent environments, Gymnasium-Robotics also includes multi-agent factorizations through MaMuJoCo, where agents cooperate or compete in factorized versions of classic MuJoCo environments. Environments such as the Maze Navigation task challenge agents like the Ant or a force-controlled ball to traverse complex mazes. Moreover, the library introduces multitask setups like the Franka Kitchen, where a 9-DoF Franka robot interacts with household items to achieve multiple objectives, and the Adroit Arm environments, where the Shadow Hand’s functionality is extended with arm movements to complete intricate tasks like hammering, door opening, or object manipulation.
The library also features a Multi-goal API, which allows agents to learn tasks with varying goals, using the GoalEnv class for environments where the agent’s objective is to match an achieved goal with a desired one. This is particularly useful for reinforcement learning techniques like Hindsight Experience Replay (HER), which can substitute goals during training to increase learning efficiency. Overall, Gymnasium-Robotics offers a rich platform for both single- and multi-agent reinforcement learning in robotic control, designed to handle complex physical interactions and flexible goal-driven tasks.
| Example observations | Gripper/joint/arm: state/position/rotation, Object position. |
| Example actions | Change arm state. |
| Example reward | achieved goal vs. desired_goal |