Skip to content

Policy Deployment

Policy Runner Base Class

Section titled “Policy Runner Base Class”

The policy runner base class PolicyRunnerBase defines input observation, output action, and inference lifecycle.

Interface Example

Section titled “Interface Example”
virtual RobotAction getRobotAction(const RobotBasicState&, const UserCommand&) = 0;
virtual void OnEnter(const RobotBasicState&) = 0;
virtual void SetDecimation(int d);

ONNX Policy Implementation

Section titled “ONNX Policy Implementation”

Lite3PolicyRunner is the ONNX policy implementation for Lite3.

Model Loading Example

Section titled “Model Loading Example”
session_ = Ort::Session(env_, policy_path_.c_str(), session_options_);

Observation and Action Dimensions

Section titled “Observation and Action Dimensions”
const int motor_num = 12;
const int observation_dim = 45;
const int action_dim = 12;

Observation vector composition:

current_observation_ << base_omega,          // 3D (angular velocity * 0.25)
                         projected_gravity,   // 3D (gravity projection)
                         command,             // 3D (velocity command)
                         joint_pos_rl,        // 12D (joint position - default position)
                         joint_vel_rl,        // 12D (joint velocity * 0.05)
                         last_action_eigen;   // 12D (previous action)

Inference Example

Section titled “Inference Example”
VecXf current_action_eigen = Onnx_infer(current_observation_);

Robot Action Output

Section titled “Robot Action Output”
for (int i = 0; i < 4; ++i){
    robot_action.goal_joint_pos.segment(i*3, 3) = tmp_action_eigen.segment(i*3, 3);
}
robot_action.kp = kp_;   // Vec3f(30, 30, 30).replicate(4, 1)
robot_action.kd = kd_;   // Vec3f(1, 1, 1).replicate(4, 1)

Joint Ordering Mapping

Section titled “Joint Ordering Mapping”

The policy and robot may have different joint orderings. Permutation indices are automatically computed via name mapping:

std::vector<std::string> robot_order = {
    "FL_HipX_joint", "FL_HipY_joint", "FL_Knee_joint",
    "FR_HipX_joint", "FR_HipY_joint", "FR_Knee_joint",
    "HL_HipX_joint", "HL_HipY_joint", "HL_Knee_joint",
    "HR_HipX_joint", "HR_HipY_joint", "HR_Knee_joint"};
robot2policy_idx = generate_permutation(robot_order, policy_order);
policy2robot_idx = generate_permutation(policy_order, robot_order);

Policy Deployment Notes

Section titled “Policy Deployment Notes”

Model File Location

Section titled “Model File Location”

The model file is located at:

src/Lite3_sdk_deploy/policy/policy.onnx

Default Posture

Section titled “Default Posture”
dof_default_eigen_policy << 0.0000, -0.8000, 1.6000,
                            0.0000, -0.8000, 1.6000,
                            0.0000, -0.8000, 1.6000,
                            0.0000, -0.8000, 1.6000;

Action Scaling

Section titled “Action Scaling”
std::vector<float> action_scale_robot = {0.125, 0.25, 0.25,
                                         0.125, 0.25, 0.25,
                                         0.125, 0.25, 0.25,
                                         0.125, 0.25, 0.25};