Skip to content

environment.md

Latest commit

 

History

History
101 lines (64 loc) · 5.07 KB

environment.md

File metadata and controls

101 lines (64 loc) · 5.07 KB

Environment

Environment is the component that interact with the robot. It receives the action from the policy and executes it on the robot.

Environment

Environment. It is an abstract class that defines the interface for the environment. The environment can be either a real robot or a simulated robot.

Environment offer feedback properties. including:

  • dof_pos: np.ndarray: the angle of the joints.
  • dof_vel: np.ndarray: the velocity of the joints.
  • base_quat: np.ndarray: the quaternion of the root. w last.
  • base_ang_vel: np.ndarray: the angular velocity of the root.
  • base_lin_acc: np.ndarray: the linear acceleration of the root.

optional variables, may not be available in all envs according to config:

  • base_pos: np.ndarray: the 3D position of the root.
  • base_lin_vel: np.ndarray: the linear velocity of the root.
  • torso_pos: np.ndarray: the 3D position of the torso.
  • torso_quat: np.ndarray: the quaternion of the torso.
  • torso_ang_vel: np.ndarray: the angular velocity of the torso.
  • fk_info: dict: the forward kinematics information:
    • {body_name}:
      • pos: np.ndarray: the 3D position of the each link.
      • quat: np.ndarray: the quaternion of the each link. w last.
      • ang_vel: np.ndarray: the angular velocity of the each link.
      • lin_vel: np.ndarray: the linear velocity of the each link.

✨ Feature: Dynamic DoF Config

The DoFConfig in environments defines joint_names, stiffness, default_pos,etc. It supports:

  • Subset extension & cropping — easily handle locked DoFs using _subset.
  • Dynamic override by policy config — policies can override environment DoF settings, even when joint_names are not aligned.
    • For instance, run a 12-DoF policy on a 29-DoF environment.
  • Strict checks — all configs are validated in DoFConfig to ensure safety.

✨ Feature: Born Place Alignment

The robot’s base quaternion and base position can be easily aligned during deployment.

This allows the robot’s starting pose to be corrected to the expected zero point, without adding complex world2init transformations inside the policy.

✨ Feature: Odometry Support

RoboJuDo supports odometry, enabling tracking of the robot’s global position and velocity. We have two options now:

  • ZED: Use a ZED camera for odometry. zed_proxy submodule is needed.
  • UNITREE: Use the built-in sport state service of Unitree robots.

This allows policies and controllers to leverage absolute positioning for more reliable deployment and motion alignment.

✨ Feature: Forward Kinematic

RoboJuDo provides simple access to rich information such as link poses, joint positions, and end-effector states. This makes it easy to implement advanced controllers and monitoring tools without additional computation overhead.

💡Tips: You can use the debug_viz option for real-time sim2real monitoring. Check ForwardKinematicCfg

debug_viz

Check Environment and EnvCfg for more details.

We provide following Environments:

Environment > DummyEnv

DummyEnv is the environment that does nothing. It is a subclass of Environment and implements the interface defined in Environment. It is mainly used for testing purposes. It will print some debug info when running.

script: dummy_env.py

Environment > MujocoEnv

MujocoEnv is the environment that simulates the robot using Mujoco. It is a subclass of Environment and implements the interface defined in Environment.

script: mujoco_env.py

Environment > UnitreeEnv

UnitreeEnv is the environment that simulates the robot using Unitree's python SDK. It is a subclass of Environment and implements the interface defined in Environment.

unitree_sdk2py is needed.

script: unitree_env.py

Environment > UnitreeCppEnv

UnitreeCppEnv is the environment that simulates the robot using Unitree's C++ SDK. It is a subclass of Environment and implements the interface defined in Environment.

script: unitree_cpp_env.py

UnitreeCpp is needed.

This environment is implemented with UnitreeCpp. In our test onboard Unitree G1, UnitreeCppEnv is faster almost 100 times than UnitreeEnv in step().