legged_robot_config.py

This document provides detailed descriptions of all configuration parameters defined in legged_robot_config.py.

---

LeggedRobotCfg

env Environment Configuration

Parameter	Type	Default	Description
num_envs	int	4096	Number of parallel environments
num_observations	int	48	Size of the observation vector
num_privileged_obs	int	None	Number of privileged observations. If not None, step() returns priviledge_obs_buf (critic obs for asymmetric training), otherwise returns None
num_actions	int	12	Size of the action vector
send_timeouts	bool	True	Whether to send timeout information to the algorithm
episode_length_s	float	20	Episode length in seconds
env_spacing	float	2.0	Spacing between environments in the scene, only for plane terrain
fail_to_terminal_time_s	float	0.1	Time before a fail state leads to environment reset (seconds)
debug	bool	False	Whether to enable debug drawings in the simulator
debug_draw_height_points_around_base	bool	False	Whether to obtain height measurements around the base
debug_draw_height_points_around_feet	bool	False	Whether to obtain height measurements around the feet (9 points around each foot)
debug_draw_terrain_height_points	bool	False	Whether to draw all height points of the terrain
debug_draw_key_body_points	bool	False	Whether to draw key body points for mimic tasks
max_projected_gravity	float	-0.1	Maximum allowed projected gravity in z-axis direction

---

terrain Terrain Configuration

Parameter	Type	Default	Description
mesh_type	str	‘plane’	Terrain mesh type: ‘plane’, ‘heightfield’, ‘trimesh’
plane_length	float	200.0	Plane size [m], default is 200x200x10
horizontal_scale	float	0.1	Distance between height samples in x and y direction [m]
vertical_scale	float	0.005	Distance between height samples in z direction [m]
border_size	float	5	Length of the border surrounding the terrain [m]
border_height	float	1.0	Height of the border surrounding the terrain [m]
curriculum	bool	False	Whether to use terrain curriculum, starting from easier terrains and gradually increasing difficulty
static_friction	float	1.0	Coefficient of static friction of the terrain
dynamic_friction	float	1.0	Coefficient of dynamic friction of the terrain
restitution	float	0.	Coefficient of restitution of the terrain
obtain_terrain_info_around_feet	bool	False	Whether to obtain terrain height information around feet (default: 9 points around each foot)
measure_heights	bool	False	Whether to obtain height measurements
measured_points_x	list	[-0.8, …, 0.8]	X positions of height sampling around the base (relative to robot base), 1m x 1.6m rectangular area
measured_points_y	list	[-0.5, …, 0.5]	Y positions of height sampling around the base (relative to robot base)
selected	bool	False	Whether to select a unique terrain type and pass all arguments
terrain_kwargs	dict	None	Dictionary of arguments for selected terrain
max_init_terrain_level	int	1	Starting curriculum level
terrain_length	float	6.0	Length of each subterrain [m] (X direction)
terrain_width	float	6.0	Width of each subterrain [m] (Y direction)
platform_size	float	3.0	Size of the flat platform at the center of each subterrain [m]
num_rows	int	4	Number of terrain rows (levels), X direction
num_cols	int	4	Number of terrain columns (types), Y direction
terrain_proportions	list	[0.1, 0.1, 0.35, 0.25, 0.2]	Terrain type proportions: [smooth slope, rough slope, stairs up, stairs down, discrete]
slope_treshold	float	0.75	Trimesh only: slopes above this threshold will be corrected to vertical surfaces

---

init_state Initial State Configuration

Parameter	Type	Default	Description
pos	list	[0.0, 0.0, 1.]	Initial position x, y, z [m]
rot	list	[0.0, 0.0, 0.0, 1.0]	Initial orientation quaternion x, y, z, w. Note: Gym format is xyzw, Genesis format is wxyz
lin_vel	list	[0.0, 0.0, 0.0]	Initial linear velocity x, y, z [m/s]
ang_vel	list	[0.0, 0.0, 0.0]	Initial angular velocity x, y, z [rad/s]
roll_random_scale	float	0.0	Roll angle randomization range
pitch_random_scale	float	0.0	Pitch angle randomization range
yaw_random_scale	float	0.0	Yaw angle randomization range
default_joint_angles	dict	{“joint_a”: 0., “joint_b”: 0.}	Target joint angles when action = 0.0

---

control Control Configuration

Parameter	Type	Default	Description
control_type	str	‘P’	Control type: ‘P’ (position), ‘V’ (velocity), ‘T’ (torque)
stiffness	dict	{‘joint_a’: 10.0, ‘joint_b’: 15.}	PD control stiffness [N*m/rad]
damping	dict	{‘joint_a’: 1.0, ‘joint_b’: 1.5}	PD control damping [Nms/rad]
action_scale	float	0.5	Action scale factor, target angle = actionScale * action + defaultAngle
dt	float	0.02	Control frequency, default 50Hz
decimation	int	4	Ratio of control action updates to policy updates

---

asset Asset Configuration

Parameter	Type	Default	Description
name	str	None	Asset name
file	str	“”	URDF file path
foot_name	str	“”	Foot body name, bodies containing this substring will be considered as feet
key_bodies	list	[]	List of important bodies to be tracked in mimic tasks
penalize_contacts_on	list	[]	Penalize contacts on links containing these substrings
terminate_after_contacts_on	list	[]	Terminate episode after contacts on links containing these substrings
fix_base_link	bool	False	Whether to fix the base link to the world frame
obtain_link_contact_states	bool	False	Whether to obtain contact states of specific links, can be used for privileged policy
contact_state_link_names	list	[“thigh”, “calf”, “foot”]	Link names for obtaining contact states
base_link_name	str	“”	Full name of the base link
self_collisions	int	0	Self-collision setting: 1 to disable, 0 to enable (bitwise filter)
dof_names	list	[“joint_a”, “joint_b”]	Sequence of DOFs in actions and observations
dof_armature	list	[0.0]	Armature of each DOF
links_to_keep	list	[]	Genesis only: links that are not merged due to fixed joints
dof_vel_limits	list	[]	DOF velocity limits [rad/s], obtained from URDF
disable_gravity	bool	False	IsaacGym/IsaacLab only: whether to disable gravity
collapse_fixed_joints	bool	True	IsaacGym/IsaacLab only: whether to merge bodies connected by fixed joints
default_dof_drive_mode	int	3	IsaacGym/IsaacLab only: DOF drive mode (0 none, 1 position target, 2 velocity target, 3 effort)
replace_cylinder_with_capsule	bool	False	IsaacGym/IsaacLab only: replace collision cylinders with capsules, improves simulation speed and stability
flip_visual_attachments	bool	False	IsaacGym/IsaacLab only: whether to flip certain .obj meshes from y-up to z-up
density	float	0.001	Density
angular_damping	float	0.	Angular damping
linear_damping	float	0.	Linear damping
max_angular_velocity	float	1000.	Maximum angular velocity
max_linear_velocity	float	1000.	Maximum linear velocity
armature	float	0.	Armature
thickness	float	0.01	Thickness

---

rewards Reward Configuration

Parameter	Type	Default	Description
only_positive_rewards	bool	True	Whether to only use positive rewards
tracking_sigma	float	0.25	Tracking reward calculation parameter: reward = exp(-error^2/sigma)
soft_dof_pos_limit	float	1.	Joint position soft limit (percentage of URDF limits), values above this limit are penalized
soft_dof_vel_limit	float	1.	Joint velocity soft limit
soft_torque_limit	float	1.	Torque soft limit
base_height_target	float	1.	Target base height
foot_clearance_target	float	0.04	Desired foot clearance above ground [m]
foot_height_offset	float	0.0	Height of the foot coordinate origin above ground [m]
foot_clearance_tracking_sigma	float	0.01	Sigma value for foot clearance tracking

scales Reward Weights

Parameter	Type	Default	Description
termination	float	-0.0	Termination penalty weight
tracking_lin_vel	float	0	Linear velocity tracking reward weight
tracking_ang_vel	float	0	Angular velocity tracking reward weight
lin_vel_z	float	0	Z-direction linear velocity penalty weight
ang_vel_xy	float	0	XY-plane angular velocity penalty weight
orientation	float	-0.	Orientation penalty weight
torques	float	0	Torque penalty weight
dof_vel	float	-0.	Joint velocity penalty weight
dof_acc	float	0	Joint acceleration penalty weight
base_height	float	-0.	Base height penalty weight
feet_air_time	float	0	Feet air time reward weight
collision	float	0	Collision penalty weight
feet_stumble	float	-0.0	Feet stumble penalty weight
action_rate	float	0	Action rate penalty weight
dof_pos_stand_still	float	-0.	Joint position penalty weight when standing still

---

commands Command Configuration

Parameter	Type	Default	Description
curriculum	bool	False	Whether to use command curriculum
max_curriculum	float	1.	Maximum curriculum level
num_commands	int	4	Number of commands, default: lin_vel_x, lin_vel_y, ang_vel_yaw, heading
resampling_time	float	10.	Time interval for command resampling [s]
heading_command	bool	True	If True: compute angular velocity command from heading error
curriculum_threshold	float	0.8	Curriculum learning threshold, increase command range when tracking reward exceeds this threshold
zero_cmd_prob	float	0.4	Probability of sampling zero command during resampling, encourages robot to learn standing still behavior

ranges Command Ranges

Parameter	Type	Default	Description
lin_vel_x	list	[-1.0, 1.0]	X-direction linear velocity range [m/s]
lin_vel_y	list	[-1.0, 1.0]	Y-direction linear velocity range [m/s]
ang_vel_yaw	list	[-1, 1]	Yaw angular velocity range [rad/s]
heading	list	[-3.14, 3.14]	Heading angle range [rad]

---

domain_rand Domain Randomization Configuration

Parameter	Type	Default	Description
randomize_friction	bool	True	Whether to randomize rigid body friction coefficient
friction_range	list	[0.5, 1.25]	Friction coefficient randomization range
randomize_base_mass	bool	True	Whether to randomize base mass
added_mass_range	list	[-1., 1.]	Added mass range [kg]
push_robots	bool	True	Whether to apply random velocity perturbations to the base
push_interval_s	float	15	Time interval for pushing robots [s]
max_push_vel_xy	float	1.	Maximum push velocity in XY plane [m/s]
randomize_com_displacement	bool	True	Whether to randomize center of mass position to simulate modeling errors
com_pos_x_range	list	[-0.01, 0.01]	CoM x-position randomization range [m]
com_pos_y_range	list	[-0.01, 0.01]	CoM y-position randomization range [m]
com_pos_z_range	list	[-0.01, 0.01]	CoM z-position randomization range [m]
randomize_ctrl_delay	bool	False	Whether to apply random delay to actions to simulate control loop latency
ctrl_delay_step_range	list	[0, 1]	Control delay step range
randomize_pd_gain	bool	False	Whether to randomize PD gains by a scale factor
kp_range	list	[0.8, 1.2]	Kp scale range
kd_range	list	[0.8, 1.2]	Kd scale range
randomize_joint_armature	bool	False	Whether to randomize joint armature (significantly slows simulation in Genesis)
joint_armature_range	list	[0.0, 0.05]	Joint armature range [Nms/rad]
randomize_joint_friction	bool	False	Whether to randomize joint friction
joint_friction_range	list	[0.0, 0.1]	Joint friction range
randomize_joint_damping	bool	False	Whether to randomize joint damping
joint_damping_range	list	[0.0, 1.0]	Joint damping range
push_links	bool	False	Whether to apply random push forces to robot links
max_push_force	float	10.0	Maximum magnitude of random push force applied to each link [N]
push_links_interval_s	float	15.0	Time interval between random pushes [s]

---

normalization Normalization Configuration

Parameter	Type	Default	Description
clip_observations	float	100.	Observation clipping range
clip_actions	float	100.	Action clipping range

obs_scales Observation Scales

Parameter	Type	Default	Description
lin_vel	float	1.0	Linear velocity observation scale factor
ang_vel	float	0.25	Angular velocity observation scale factor
dof_pos	float	1.0	Joint position observation scale factor
dof_vel	float	0.05	Joint velocity observation scale factor
height_measurements	float	5.0	Height measurement observation scale factor

---

noise Noise Configuration

Parameter	Type	Default	Description
add_noise	bool	True	Whether to add noise
noise_level	float	1.0	Noise scale factor

noise_scales Noise Scales

Parameter	Type	Default	Description
dof_pos	float	0.01	Joint position noise scale
dof_vel	float	0.5	Joint velocity noise scale
lin_vel	float	0.1	Linear velocity noise scale
ang_vel	float	0.2	Angular velocity noise scale
gravity	float	0.05	Gravity observation noise scale
height_measurements	float	0.1	Height measurement noise scale

---

constraints Constraints Configuration

Used for CaT (Constraints as Termination) method.

Parameter	Type	Default	Description
limits	class	-	Constraint limits configuration

---

viewer Viewer Configuration

Parameter	Type	Default	Description
ref_env	int	0	Reference environment index
pos	list	[4.0, 4.0, 2.0]	Camera position relative to robot position [m]
lookat	list	[0., 0, 0.]	Point the camera looks at (relative to robot position) [m]
rendered_envs_idx	list	[0, 1, 2, 3, 4]	Genesis only: list of environment indices to render

---

sensor Sensor Configuration

Parameter	Type	Default	Description
add_depth	bool	False	Whether to add depth camera
use_warp	bool	False	Whether to use warp-based model

depth_camera_config Depth Camera Configuration

Parameter	Type	Default	Description
num_sensors	int	1	Number of sensors
num_history	int	1	History frames for depth images
near_clip	float	0.1	Near clipping plane
far_clip	float	10.0	Far clipping plane
near_plane	float	0.1	Near plane
far_plane	float	10.0	Far plane
resolution	tuple	(80, 60)	Image resolution
horizontal_fov_deg	float	75	Horizontal field of view [degrees]
pos	tuple	(0.3, 0.0, 0.1)	Camera position
euler	tuple	(0.0, 0.0, 0.0)	Camera Euler angles
decimation	int	5	Decimation factor
calculate_depth	bool	True	Warp only: whether to calculate depth
segmentation_camera	bool	False	Warp only: whether to use segmentation camera
return_pointcloud	bool	False	Warp only: whether to return point cloud
pointcloud_in_world_frame	bool	False	Warp only: whether point cloud is in world frame

---

sim Simulation Configuration

Parameter	Type	Default	Description
dt	float	0.005	Simulation timestep, default 200Hz
substeps	int	1	Number of substeps
max_collision_pairs	int	100	Genesis only: maximum number of collision pairs, more collision pairs will occupy more GPU memory and slow down simulation
IK_max_targets	int	2	Genesis only: number of IK targets, fewer targets will reduce memory usage
gravity	list	[0., 0., -9.81]	IsaacGym only: gravity acceleration [m/s^2]
up_axis	int	1	IsaacGym only: up axis direction, 0 for y, 1 for z
use_gpu_pipeline	bool	True	IsaacGym only: whether to use GPU pipeline

physx PhysX Engine Configuration (IsaacGym only)

Parameter	Type	Default	Description
use_gpu	bool	True	Whether to use GPU
num_subscenes	int	0	Number of subscenes
num_threads	int	10	Number of threads
solver_type	int	1	Solver type: 0 for pgs, 1 for tgs
num_position_iterations	int	4	Number of position iterations
num_velocity_iterations	int	0	Number of velocity iterations
contact_offset	float	0.01	Contact offset [m]
rest_offset	float	0.0	Rest offset [m]
bounce_threshold_velocity	float	0.5	Bounce threshold velocity [m/s]
max_depenetration_velocity	float	1.0	Maximum depenetration velocity
max_gpu_contact_pairs	int	2**23	Maximum GPU contact pairs
default_buffer_size_multiplier	int	5	Default buffer size multiplier
contact_collection	int	2	Contact collection mode: 0 never, 1 last substep, 2 all substeps

---

LeggedRobotCfgPPO

PPO algorithm configuration class.

Basic Parameters

Parameter	Type	Default	Description
seed	int	1	Random seed
runner_class_name	str	‘OnPolicyRunner’	Runner class name

---

policy Policy Configuration

Parameter	Type	Default	Description
clip_actions	float	LeggedRobotCfg.normalization.clip_actions	Action clipping range
init_noise_std	float	1.0	Initial noise standard deviation
actor_hidden_dims	list	[512, 256, 128]	Actor network hidden layer dimensions
critic_hidden_dims	list	[512, 256, 128]	Critic network hidden layer dimensions
activation	str	‘elu’	Activation function, options: elu, relu, selu, crelu, lrelu, tanh, sigmoid

RNN Configuration (for ActorCriticRecurrent only)

Parameter	Type	Default	Description
rnn_type	str	‘lstm’	RNN type
rnn_hidden_size	int	512	RNN hidden layer size
rnn_num_layers	int	1	Number of RNN layers

---

algorithm Algorithm Configuration

Parameter	Type	Default	Description
value_loss_coef	float	1.0	Value loss coefficient
use_clipped_value_loss	bool	True	Whether to use clipped value loss
clip_param	float	0.2	PPO clipping parameter
entropy_coef	float	0.01	Entropy coefficient
num_learning_epochs	int	5	Number of learning epochs
num_mini_batches	int	4	Number of mini batches, mini batch size = num_envs * nsteps / nminibatches
learning_rate	float	1.e-3	Learning rate
schedule	str	‘adaptive’	Learning rate schedule: ‘adaptive’ or ‘fixed’
gamma	float	0.99	Discount factor
lam	float	0.95	GAE lambda parameter
desired_kl	float	0.01	Target KL divergence
max_grad_norm	float	1.	Maximum gradient clipping norm
use_spo	bool	False	Whether to use SPO (Simple Policy Optimization). Note: SPO may be incompatible with default PPO parameters, recommended: learning_rate=2.5e-4, schedule=’fixed’

---

runner Runner Configuration

Parameter	Type	Default	Description
policy_class_name	str	‘ActorCritic’	Policy class name
algorithm_class_name	str	‘PPO’	Algorithm class name
num_steps_per_env	int	24	Number of steps per environment per iteration
max_iterations	int	1500	Maximum number of policy update iterations
sync_wandb	bool	False	Whether to sync logs to wandb
save_interval	int	50	Check for potential saves every this many iterations
experiment_name	str	‘test’	Experiment name
run_name	str	‘’	Run name
resume	bool	False	Whether to resume from checkpoint
load_run	int	-1	Run ID to load, -1 for last run
checkpoint	int	-1	Checkpoint ID, -1 for last saved model
resume_path	str	None	Resume path, updated from load_run and checkpoint