Parameter Tuning Guide¶

Every tunable parameter for G1 locomotion training. Two files control everything.

File 1: Environment Config¶

Path: source/unitree_rl_lab/unitree_rl_lab/tasks/locomotion/robots/g1/29dof/velocity_env_cfg.py

This file defines the robot, terrain, observations, rewards, and termination conditions.

Reward Weights¶

These are the most impactful parameters. Positive weights encourage behavior, negative weights penalize it.

Task rewards (what you want the robot to do):

Penalties (what you want the robot to avoid):

Tuning tips:

• Increase track_lin_vel_xy to make the robot prioritize speed over stability
• Increase flat_orientation_l2 to make it more conservative about tilting
• Increase alive to make survival the top priority (useful early in a new config)
• Increase gait to enforce a more rhythmic stepping pattern
• Decrease joint_deviation_arms (closer to 0) to let arms move more freely

Velocity Command Ranges¶

Controls what velocities are sampled during training:

# Initial training range (starts conservative, widens via curriculum)
lin_vel_x=(-0.1, 0.1)    # forward/backward (m/s)
lin_vel_y=(-0.1, 0.1)    # lateral (m/s)
ang_vel_z=(-0.1, 0.1)    # yaw rate (rad/s)

# Maximum range (curriculum expands to this)
lin_vel_x=(-0.5, 1.0)    # max: 1 m/s forward, 0.5 m/s backward
lin_vel_y=(-0.3, 0.3)    # max: 0.3 m/s lateral
ang_vel_z=(-0.2, 0.2)    # max: 0.2 rad/s turning

Change limit_ranges to train for faster walking or sharper turns.

Domain Randomization (Events)¶

Randomizations applied during training to improve robustness:

Tuning tips:

• Increase push_robot velocity range to train a more stable robot
• Widen add_base_mass range if the robot will carry objects
• Increase static_friction_range minimum to avoid training on very slippery surfaces

Termination Conditions¶

When an episode ends:

Lower bad_orientation limit to terminate earlier when tilting (stricter). Raise it to let the robot recover from bigger tilts.

Terrain¶

Currently set to flat terrain:

sub_terrains={
    "flat": terrain_gen.MeshPlaneTerrainCfg(proportion=0.5),
}

Can be changed to include slopes, stairs, or rough terrain for more challenging training.

Simulation Parameters¶

decimation = 4          # policy runs every 4 physics steps
episode_length_s = 20.0 # 20-second episodes
sim.dt = 0.005          # physics timestep (200 Hz)
# Policy frequency = 200 / 4 = 50 Hz

Observations¶

What the policy sees (in order):

History length: 5 timesteps (all observations stacked across 5 steps, creating the 480-dim input)

Action Space¶

JointPositionAction = mdp.JointPositionActionCfg(
    asset_name="robot",
    joint_names=[".*"],    # all 29 joints
    scale=0.25,            # output scaled by 0.25
    use_default_offset=True  # actions are offsets from default pose
)

scale=0.25 means a policy output of 1.0 moves the joint 0.25 radians from default. Increase for more aggressive movements.

File 2: Training Algorithm Config¶

Path: source/unitree_rl_lab/unitree_rl_lab/tasks/locomotion/agents/rsl_rl_ppo_cfg.py

This file controls the neural network architecture and learning hyperparameters.

Network Architecture¶

actor_hidden_dims=[512, 256, 128]   # policy network layers
critic_hidden_dims=[512, 256, 128]  # value function layers
activation="elu"                     # activation function
init_noise_std=1.0                   # initial exploration noise

Larger networks learn more complex behaviors but train slower. The current 512-256-128 architecture is standard for locomotion.

Learning Hyperparameters¶

Training Loop¶

Tuning tips:

• Lower learning_rate (e.g. 5e-4) for more stable but slower training
• Increase entropy_coef if the robot gets stuck in a local optimum (e.g. standing still)
• Increase num_learning_epochs to squeeze more learning from each batch of experience
• gamma=0.99 means the robot cares about reward up to ~100 steps in the future

Summary: What to Change Where¶