| Definition | Panda-Gym is a flexible simulation environment for the Panda robotical arm, offering six manipulation tasks with varying control modes and reward structures for testing reinforcement learning algorithms. |
| No of environments | 34 |
Panda-Gym is a versatile environment designed for simulating the Franka Emika Panda robot, featuring six distinct tasks that challenge its manipulation capabilities. The tasks include Reach, where the robot places its end-effector at a target position; Push, which involves pushing a cube to a designated spot; Slide, where the robot slides an object to a target; Pick and Place, requiring the robot to pick up and position an object; Stack, which tasks the robot with stacking two cubes; and Flip, where it must flip a cube to achieve a specific orientation. These tasks can be approached using two control modes: end-effector displacement control, which directs the robot’s end-effector movement, and joint control, which manages the individual joints of the robot.
The environment offers both sparse and dense reward structures, allowing for flexibility in training approaches. In the default setting, rewards are sparse, providing a reward only upon task completion, while dense rewards give incremental feedback based on the robot’s proximity to task completion. This allows users to tailor the training experience according to their needs, whether by using the default sparse reward with end-effector control or selecting from a range of dense or joint control variations. Each task and reward combination is available, making Panda-Gym a comprehensive testbed for evaluating various reinforcement learning algorithms in robotic manipulation scenarios.
| Example observations | |
| Example actions | |
| Example reward |