This repository implements a 2D or 3D robotic arm environment for reinforcement learning, featuring a highly scalable action space. The goal is to control the arm to reach a target position in the environment.

The agent receives an observation of the environment, which includes the target position, the current end effector position, and the current joint angles of the robot arm. The agent then takes an action to change the joint angles, and receives a reward based on the distance between the end effector and the target position.

This environment is designed to be adaptable to various scenarios by offering a scalable action space. Whether you're dealing with a simple 2-joint arm or a complex multi-jointed manipulator, this environment can accommodate your needs.

The required libraries can be seen here. You can always install the manually but we recommend to install them with poetry.

poetry install

Import either the discrete or continuous version and a task you want to complete.

from ik_rl.environment import InvKinDiscrete, InvKinEnvContinous
from ik_rl.task import ReachGoalTask

Create an environment and task instance:

task = ReachGoalTask(arm_reach=n_joints * segment_length)
env = InvKinEnvContinous(task=task, robot_config={"n_joints": n_joints})

You are always free to replace ReachGoalTask with the specific task you want the robot arm to complete. This class should inherit from ik_rl.task.base_task.BaseTask.

Reset the environment:

observation = env.reset()

Take an action and get the next observation, reward, and done flag: Python

action = env.action_space.sample()
observation, reward, done, info = env.step(action)

Render the environment (optional):

env.render()

Close the environment:

env.close()

For an working example please refer to the example notebook. For the example you are required to install stable baselines 3.

• InvKinEnv: This is the base class for all inverse kinematic environments. It supports both 2D and 3D robot arms, and allows for choosing between discrete and continuous action spaces.
• InvKinDiscrete: This class provides a discrete action space, where the agent can choose from a set of predefined joint angle changes.
• InvKinEnvContinous: This class provides a continuous action space, where the agent can specify the desired joint angle changes directly.

We welcome contributions to this project! Please see the CONTRIBUTING.md file for more information. Planned features are:

• expand to 3-D (Optional n-D)
• add obstacles to env
  • polygons, circles, ...
  • in 3D add table top (z has to be >= 0) constraint
  • method to compute if is possible to solve for a given target position given the env with obstacles
• add parallel computing for as many arms as I want

This project has benefited from the contributions of several individuals and resources. We would like to express our sincere gratitude to:

The developers of the following libraries:

• OpenAI Gym
• Farama Gymnasium
• NumPy
• Matplotlib
• PyTorch (optional)

The contributors to various online communities and forums who shared their knowledge and insights on reinforcement learning and inverse kinematics.
Jasper Hoffmann for their valuable feedback, guidance, and support during the development process.
If you have contributed to this project and are not mentioned here, please feel free to reach out and we will be happy to include your acknowledgment.

We are grateful for the support of the community and look forward to further contributions to this project.