Skip to content

policy.md

Latest commit

 

History

History
204 lines (134 loc) · 9.77 KB

policy.md

File metadata and controls

204 lines (134 loc) · 9.77 KB

Policy

Policy is the component that controls the robot. It receives the env_data from the environment, ctrl_data from the controller, organize the observation and infer the action fo robot.

Policy

Policy is the base class for all policies. It defines the interface for the policy, as in base_policy.py

We provide the following policies:

Policy > UnitreePolicy

UnitreePolicy is the policy that controls the robot using the Unitree official policy.

script: unitree_policy.py

To control the robot using UnitreePolicy, you can refer _get_commands():

commands:

  • commands[0], [-1, 1], control the robot to walk forward and backward
  • commands[1], [-1, 1], control the robot to walk left and right
  • commands[2], [-1, 1], control the robot to turn left and right

for instance, use JoystickCtrl to control:

def _get_commands(self, ctrl_data: dict) -> list[float]:
    commands = np.zeros(3)
    for key in ctrl_data.keys():
        if key in ["JoystickCtrl", "UnitreeCtrl"]:
            axes = ctrl_data[key]["axes"]
            lx, ly, rx, ry = axes["LeftX"], axes["LeftY"], axes["RightX"], axes["RightY"]

            commands[0] = command_remap(ly, self.commands_map[0])
            commands[1] = command_remap(lx, self.commands_map[1])
            commands[2] = command_remap(rx, self.commands_map[2])
            break
    return commands

UnitreeWoGaitPolicy

For Unitree G1, we also provide UnitreeWoGaitPolicy, which supports the new Unitree-G1-29dof-Velocity conig from unitree_rl_lab.

The difference is that UnitreeWoGaitPolicy does not include gait in the observation, so the robot will not keep stepping when standing.

script: unitree_policy.py

Policy > AMOPolicy

AMOPolicy is the policy that controls the robot using the AMO.

script: amo_policy.py

To control the robot using AMOPolicy, you can refer _get_commands():

commands:

  • commands[0], [-1, 1], control the robot to walk forward and backward
  • commands[1], [-1, 1], control the robot to turn left and right
  • commands[2], [-1, 1], control the robot to walk left and right
  • commands[3], [-0.5, 0.8], control the robot torso height
  • commands[4], [-1.57, 1.57], control the robot torso yaw
  • commands[5], [-0.52, 1.57], control the robot torso pitch
  • commands[6], [-0.7, 0.7], control the robot torso roll

You can apply your own controller to control the robot using AMOPolicy. Just set the commands in _get_comands()

Policy > H2HStudentPolicy

H2HStudentPolicy is the policy that controls the robot using the human2humanoid.

PHC Submodule is needed for motionlib control. check README#setup.

script: h2hstudent_policy.py

H2HStudentPolicy is controlled by MotionH2HCtrl. check code motion_h2h_ctrl.py.

For motion source:

  • Unitree H1: Simply use the motion retargeting pipeline from human2humanoid.
  • Unitree G1: As not officially supported, we use the PHC pipeline. Our submodule patch enables 29dof G1. Check unitree_g1_29dof_fitting.yaml.

You can refer to g1_h2h config in g1_cfg.py for more details.

Policy > HugWBCPolicy

HugWBCPolicy is the policy that controls the robot using the HugWBC.

script: hugwbc_policy.py

🥺Will release soon.

Policy > BeyondMimicPolicy

BeyondMimicPolicy is the policy that controls the robot using the whole_body_tracking.

We support both G1FlatEnvCfg and G1FlatWoStateEstimationEnvCfg. For motion source, you could use the motion inside onnx policy, or use BeyondmimicCtrl with npz files.

script: beyondmimic_policy.py

BeyondMimicPolicyCfg: check example at G1BeyondMimicPolicyCfg:

  • policy_name: The name of the policy. We provive Jump_wose for test. You should put your policy in assets/models/g1/beyondmimic
  • without_state_estimator: Weather policy is WoStateEstimation. Default is True.
  • use_modelmeta_config: Whether to use modelmeta config. Default is True. If False, the policy will use config in your BeyondMimicPolicyCfg.
  • use_motion_from_model: Whether to use motion in the onnx model. Default is True. If False, you need to enable BeyondMimicCtrl. .

You can refer to g1_beyondmimic and g1_beyondmimic_with_ctrl in g1_cfg.py for details.

Policy > AsapPolicy

AsapPolicy is the policy that controls the robot using the ASAP.

Also, KungfuBot is supported by AsapPolicy.

RoboJuDo support both deepmimic and decoupled_locomotion of the official repo, implemeted as AsapPolicy and AsapLocoPolicy.

We fully reproduced the original repository, including keyboard and joystick mapping:

  • i to make the robot the initial position
  • o to emergence stop the robot

for locomotion policy:

  • = to switch between tapping and walking for the locomotion policy
  • w/a/s/d to control the linear velocity
  • q/e to control the angular velocity
  • z to set all commands to zero

for policy switch:

  • [ to switch to MotionMimic
  • ] to switch to LocoMotion
  • ; toggle next mimic policy
  • ' toggle prev mimic policy

or with joystick:

  • Left to switch between tapping and walking for the locomotion policy
  • Up/Down to control the height
  • left axes to control the linear velocity
  • right axes to control the angular velocity
  • Select/Back to switch to LocoMotion
  • Start to switch to MotionMimic
  • R1/RB to toggle next mimic policy
  • L1/LB to toggle prev mimic policy

script: asap_policy.py

For your convenience, CR7_level1 checkpoint is included, you can ran sim2sim with g1_asap config in g1_asap_cfg.py.

You can add more models to assets/models/g1/asap/mimic. Any model in the official repo ASAP-sim2real, PBHC and RoboMimic_Deploy should work.

This example highlights the advantages of RoboJudo:

  • Modular code & config with easy implementation and strong readability
  • Flexible policy switching, with interpolation support.
  • Convenient external controller processing

You can refer to g1_asap and g1_asap_loco config in g1_asap_cfg.py for test and details.

Policy > KungfuBotGeneralPolicy

KungfuBotGeneralPolicy is the policy that controls the robot using the PBHC-KungfuBot2.

To be noted, this is for KungfuBot2 general model, for KungfuBot, please use AsapPolicy.

PHC Submodule is needed for motionlib control. check README.

script: kungfubot_policy.py

  • KungfuBotGeneralPolicy is controlled by MotionKungfuBotCtrl
    • check code motion_kungfubot_ctrl.py.
    • motions from PBHC pipeline are supported. Put your motion files in assets/motions/g1/phc/kungfubot/.

You can refer to g1_kungfubot2 config in g1_cfg.py for test and details.

Policy > TwistPolicy

TwistPolicy is the policy that controls the robot using the TWIST.

script: twist_policy.py

For TwistPolicy, we implement two motion source controllers:

  • TwistRedisCtrl at twist_redis_ctrl.py:

  • MotionTwistCtrl at motion_twist_ctrl.py:

    • get motion from local .pkl files.
    • this is base on the PHC MotionLib, and uses the same motion format as PHC.
      • PHC Submodule is needed. check README.
    • motions from PBHC pipeline is supported, put your motion files in assets/motions/g1/phc/.

You can refer to g1_twist config in g1_cfg.py for test and details.