unilab.tasks.locomotion.common.gait_terms

mjlab-aligned gait-quality terms shared by locomotion families.

These are the window-form feet_air_time and the clearance / swing-height / slip / angular-momentum reward terms that mjlab ships in its velocity task, plus the privileged per-foot observation terms (foot_height / foot_air_time / foot_contact / foot_contact_forces) its velocity critic consumes. mjlab reads feet through a ContactSensor (primaries with per-slot any-reduce) and a TerrainHeightSensor; the UniLab port binds one named contact sensor group per foot through the SensorTermBase cold-path contract (any sensor in the group above the threshold means foot contact, mirroring the slot reduce) and reads foot heights/velocities from body_link_pos_w / body_link_lin_vel_w through asset_cfg.body_ids, which is exact on flat terrain without raycast.

Classes

angular_momentum_penalty

Penalize whole-body angular momentum read from a named vec3 sensor.

feet_air_time

Count feet whose current air time sits inside the rewarded window.

feet_clearance

Penalize clearance-height deviation weighted by planar foot velocity.

feet_slip

Penalize foot sliding: squared xy velocity while the foot is in contact.

feet_swing_height

Penalize peak-swing-height deviation from the target, evaluated at landing.

foot_air_time

Current per-foot air time in seconds, shape (num_envs, num_feet).

foot_contact

Per-foot contact flag as float, shape (num_envs, num_feet).

foot_contact_forces

Log-compressed per-foot contact forces, shape (num_envs, 3 * num_feet).

foot_height

Per-foot vertical clearance above the terrain, shape (num_envs, num_feet).

self_collision_cost

Count monitored self-collision geom pairs reporting contact.
class unilab.tasks.locomotion.common.gait_terms.angular_momentum_penalty[source]

Bases: SensorTermBase

Penalize whole-body angular momentum read from a named vec3 sensor.

Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

__init__(cfg, env)[source]
Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

property tensor_sensor_names: tuple[str, ...]
class unilab.tasks.locomotion.common.gait_terms.feet_air_time[source]

Bases: _FootContactTerm

Count feet whose current air time sits inside the rewarded window.

Mirrors mjlab feet_air_time: air time accumulates env.step_dt while the foot has no contact and resets to zero on contact; a foot scores while threshold_min < air_time < threshold_max. When command_name is set the count is gated by the command gate.

Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

__init__(cfg, env)[source]
Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

reset(env_ids)[source]

Row boundary for stateless sensor terms.

Parameters:

env_ids (Tensor)

Return type:

None

class unilab.tasks.locomotion.common.gait_terms.feet_clearance[source]

Bases: ManagerTermBase

Penalize clearance-height deviation weighted by planar foot velocity.

Mirrors mjlab feet_clearance with foot heights taken from world-frame foot body z (exact on flat terrain) instead of a terrain height sensor.

Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

__init__(cfg, env)[source]
Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

property tensor_body_names: tuple[str, ...]
property entity_name: str
class unilab.tasks.locomotion.common.gait_terms.feet_slip[source]

Bases: _FootContactTerm

Penalize foot sliding: squared xy velocity while the foot is in contact.

Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

__init__(cfg, env)[source]
Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

class unilab.tasks.locomotion.common.gait_terms.feet_swing_height[source]

Bases: _FootContactTerm

Penalize peak-swing-height deviation from the target, evaluated at landing.

Mirrors mjlab feet_swing_height: the peak foot height accumulates while the foot is in the air; on first contact the squared relative error (peak / target_height - 1)^2 is charged and the peak resets.

Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

__init__(cfg, env)[source]
Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

reset(env_ids)[source]

Row boundary for stateless sensor terms.

Parameters:

env_ids (Tensor)

Return type:

None

class unilab.tasks.locomotion.common.gait_terms.foot_air_time[source]

Bases: _FootContactTerm

Current per-foot air time in seconds, shape (num_envs, num_feet).

Mirrors mjlab foot_air_time (ContactSensor.data.current_air_time): the timer accumulates env.step_dt while the foot has no contact and resets to zero on contact.

Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

__init__(cfg, env)[source]
Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

reset(env_ids)[source]

Row boundary for stateless sensor terms.

Parameters:

env_ids (Tensor)

Return type:

None

class unilab.tasks.locomotion.common.gait_terms.foot_contact[source]

Bases: _FootContactTerm

Per-foot contact flag as float, shape (num_envs, num_feet).

Mirrors mjlab foot_contact (ContactSensor.data.found > 0).

Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

class unilab.tasks.locomotion.common.gait_terms.foot_contact_forces[source]

Bases: _FootContactTerm

Log-compressed per-foot contact forces, shape (num_envs, 3 * num_feet).

Mirrors mjlab foot_contact_forces: the per-foot 3-D contact forces are flattened in sensor-group order and compressed as sign(f) * log1p(|f|).

Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

__init__(cfg, env)[source]
Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

class unilab.tasks.locomotion.common.gait_terms.foot_height[source]

Bases: ManagerTermBase

Per-foot vertical clearance above the terrain, shape (num_envs, num_feet).

mjlab reads this from a TerrainHeightSensor ring around each foot; the UniLab port returns the world-frame foot body z, which is identical on flat terrain.

Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

__init__(cfg, env)[source]
Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)

property tensor_body_names: tuple[str, ...]
property entity_name: str
class unilab.tasks.locomotion.common.gait_terms.self_collision_cost[source]

Bases: _FootContactTerm

Count monitored self-collision geom pairs reporting contact.

Mirrors mjlab self_collision_cost in its found-count form: each sensor group binds one contact sensor watching a geom pair, and the cost is the number of pairs with an active contact. mjlab’s force_threshold only applies when the sensor records force history, which the upstream config does not enable, so it is not ported.

Parameters:

  • cfg (ManagerTermBaseCfg)

  • env (ManagerBasedRlEnv)