dm_robotics is a collection of libraries, tools and objects used in robotics research at DeepMind.

dm_robotics is made up of a number of Python and C++ packages, including:

• AgentFlow - Python library for composing reinforcement learning (RL) agents and breaking down complex tasks into subtasks, simplifying a common challenge in robotics research which requires training RL modules for different skills and then composing them together.
• Controllers - Collection of robot controllers and action mappers, including a Cartesian 6D velocity to joint velocity mapper with collision avoidance. Written in C++, Python bindings included.
• Geometry - Python primitives for dealing with scene and robot geometry.
• Manipulation - Python utilities for generating parametric objects using the OnShape Python API and example object meshes designed to offer different affordances for manipulation tasks.
• Modular Manipulation (MoMa) - Python library for building manipulation environments, both in simulation and on real robots, which builds on DeepMind’s Composer library.
• Transformations - Python library for rigid-body transformations, including velocities and forces, aimed at providing simplicity and comprehensiveness across all canonical representations (euler, axis-angle, quaternion, homogeneous matrices).
• Vision - Python library that provides functions and ROS nodes for colour-based blob detection and 3D triangulation.