| 1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374 |
- experiment:
- name: walk_ppo
- robot:
- model_name: guguji
- urdf_path: guguji_ros2_ws/src/guguji_ros2/urdf/guguji.urdf
- joint_names:
- - left_hip_pitch_joint
- - left_knee_pitch_joint
- - left_ankle_pitch_joint
- - left_ankle_joint
- - right_hip_pitch_joint
- - right_knee_pitch_joint
- - right_ankle_pitch_joint
- - right_ankle_joint
- command_topic_prefix: /guguji/command
- ros:
- joint_state_topic: /joint_states
- tf_topic: /tf
- clock_topic: /clock
- world_control_service: /world/default/control
- sim:
- world_name: default
- # 走路训练也建议采用 service_step，便于稳定做 step / reward 对齐。
- step_mode: service_step
- control_dt: 0.05
- service_step_iterations: 50
- reset_settle_seconds: 1.2
- action_publish_delay: 0.01
- post_step_wait_seconds: 0.01
- # 训练时建议 Gazebo 以 pause 模式启动，再由训练程序按步推进仿真。
- launch_hint: ros2 launch guguji_ros2 gazebo.launch.py gui:=false pause:=true
- task:
- target_forward_velocity: 0.25
- target_base_height: null
- max_roll_rad: 0.65
- max_pitch_rad: 0.65
- min_base_height: 0.12
- rewards:
- alive_bonus: 1.0
- velocity_tracking_scale: 3.0
- velocity_tracking_sigma: 0.25
- upright_scale: 1.5
- height_scale: 0.8
- action_rate_penalty_scale: 0.04
- joint_limit_penalty_scale: 0.05
- lateral_velocity_penalty_scale: 0.15
- fall_penalty: -20.0
- training:
- algorithm: ppo
- total_timesteps: 500000
- max_episode_steps: 500
- seed: 42
- device: auto
- learning_rate: 0.0003
- n_steps: 1024
- batch_size: 256
- gamma: 0.99
- gae_lambda: 0.95
- clip_range: 0.2
- ent_coef: 0.0
- vf_coef: 0.5
- policy_net_arch: [256, 256]
- checkpoint_freq: 50000
- output_root: guguji_rl/outputs
- evaluation:
- episodes: 3
- deterministic: true
|
