experiment:
  name: balance_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
  # 强化学习训练更适合用 service_step，能让每一步更可控、更容易复现。
  step_mode: service_step
  control_dt: 0.05
  service_step_iterations: 50
  reset_settle_seconds: 1.0
  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: 0.0
  target_base_height: null
  max_roll_rad: 0.60
  max_pitch_rad: 0.60
  min_base_height: 0.14

rewards:
  alive_bonus: 1.0
  velocity_tracking_scale: 1.5
  velocity_tracking_sigma: 0.30
  upright_scale: 2.0
  height_scale: 1.0
  action_rate_penalty_scale: 0.03
  joint_limit_penalty_scale: 0.05
  lateral_velocity_penalty_scale: 0.10
  fall_penalty: -15.0

training:
  algorithm: ppo
  total_timesteps: 200000
  max_episode_steps: 400
  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: 20000
  output_root: guguji_rl/outputs

evaluation:
  episodes: 3
  deterministic: true