Ant

superdex_gym/Ant-v0 · superdex.lab.gym.envs.benchmarks.ant_env.AntEnv

Description

Ant provides quadrupedal locomotion. The reward favors forward progress and penalizes control effort and contact-wrench magnitude.

Action Space

The action is an (8,) Box. The element control is bounded [-1.0, 1.0] and applied as external forces on DOFs 6-13 with a force scale of 150.

Observation Space

The observation is an (81,) Box with the default configuration, flattened in alphabetical key order:

Key	Shape	Meaning
contact_forces	54	Six contact-wrench components for each of 9 bodies. Omitted when include_contact_in_observation=False
pose	13	num_dofs (14) - 2 excluded positions, +1 for the quaternion representation
vel	14	One per DOF

Turning off exclude_current_positions_from_observation adds 2, turning off include_contact_in_observation removes 54, and setting use_rotation_vector=True removes 1.

Rewards

The scalar reward is the sum of four terms:

Term	Value
forward	x_velocity * forward_reward_weight
ctrl	-control_cost_weight * dot(control, control)
contact	-contact_cost_weight * squared norm of the clipped contact wrench
survive	healthy_reward while healthy, else 0

Starting State

The scene is restored to its captured initial state, then uniform reset noise scaled by reset_noise_scale (default 0.1) is added to the pose and velocity.

Episode End

The episode terminates (terminated_reason = "Unhealthy") when the torso height leaves healthy_y_range, unless terminate_when_unhealthy=False. It truncates at steps_per_episode.

Arguments

In addition to the shared MochiEnvCfg fields (see the Authoring guide's base configuration section), AntEnvCfg accepts:

Field	Default	Meaning
control_frequency	20	Control frequency [Hz]
simulation_frequency	100	Simulation frequency [Hz]; 5 substeps per control step
steps_per_episode	1000	Truncation limit
reset_noise_scale	0.1	Scale of the uniform reset noise
forward_reward_weight	1	Weight of the forward-progress term
control_cost_weight	0.5	Weight of the control-cost penalty
contact_cost_weight	5e-4	Weight of the contact-cost penalty
healthy_reward	1.0	Per-step bonus while healthy
terminate_when_unhealthy	True	Terminate when the torso leaves the healthy range
healthy_y_range	(0.2, 1.0)	Torso height band [m]
contact_force_range	(-1.0, 1.0)	Per-component clip range for each body's contact wrench
exclude_current_positions_from_observation	True	Omit x and z, keeping policies translation-invariant
include_contact_in_observation	True	Include the contact-wrench observation
use_rotation_vector	False	Represent rotations with a rotation vector

Scene identity

uid_fields = (use_damping, use_gravity, use_low_friction, use_high_friction).

Variants

AntFullObservation

superdex_gym/AntFullObservation-v0

Field	Value
exclude_current_positions_from_observation	False

AntNoContact

superdex_gym/AntNoContact-v0

Field	Value
include_contact_in_observation	False

AntRotationVector

superdex_gym/AntRotationVector-v0

Field	Value
use_rotation_vector	True