Implementation:Facebookresearch Habitat lab Sensor Measure Action base classes

Knowledge Sources	Habitat-Lab
Domains	Software_Architecture, Embodied_AI
Last Updated	2026-02-15 02:00 GMT

Overview

Concrete base classes for defining custom sensors, measures, and actions in the Habitat task framework, provided by habitat-lab core.

Description

Habitat provides three base classes for extension:

• Sensor (in habitat.core.simulator): Subclass and override _get_uuid, _get_sensor_type, _get_observation_space, and get_observation
• Measure (in habitat.core.embodied_task): Subclass and override _get_uuid, reset_metric, and update_metric
• SimulatorTaskAction (in habitat.core.embodied_task): Subclass and override step to define action effects

Usage

Import the appropriate base class, create a subclass with required method overrides, then register it with the Habitat registry (next step).

Code Reference

Source Location

• Repository: habitat-lab
• File: habitat-lab/habitat/core/simulator.py (Sensor: L74-111), habitat-lab/habitat/core/embodied_task.py (Measure: L83-127, SimulatorTaskAction: L60-81)

Signature

# Sensor base class
class Sensor:
    uuid: str
    config: DictConfig
    sensor_type: SensorTypes

    def __init__(self, *args, config, sim, **kwargs):
        self.uuid = self._get_uuid(*args, **kwargs)
        self.config = config
        self.sensor_type = self._get_sensor_type(*args, **kwargs)

    def _get_uuid(self, *args, **kwargs) -> str: ...
    def _get_sensor_type(self, *args, **kwargs) -> SensorTypes: ...
    def _get_observation_space(self, *args, **kwargs) -> spaces.Space: ...
    def get_observation(self, *args, **kwargs) -> Any: ...

# Measure base class
class Measure:
    _metric: Any

    def __init__(self, *args, config, **kwargs):
        self._metric = None

    def _get_uuid(self, *args, **kwargs) -> str: ...
    def reset_metric(self, *args, **kwargs) -> None: ...
    def update_metric(self, *args, **kwargs) -> None: ...

# Action base class
class SimulatorTaskAction:
    def __init__(self, *args, config, sim, **kwargs):
        self._config = config
        self._sim = sim

    def step(self, *args, **kwargs):
        """Execute action in simulator."""
        ...

Import

from habitat.core.simulator import Sensor, SensorTypes
from habitat.core.embodied_task import Measure, SimulatorTaskAction

I/O Contract

Inputs

Name	Type	Required	Description
sim	Simulator	Yes	Simulator instance (provides scene state)
config	DictConfig	Yes	Component-specific configuration
episode	Episode	Yes	Current episode (for Sensors/Measures)
task	EmbodiedTask	Yes	Task instance (for Measures)

Outputs

Name	Type	Description
Sensor.get_observation	Any	Observation value (ndarray, scalar, etc.)
Measure._metric	Any	Metric value (scalar, dict)
SimulatorTaskAction.step	Observations	New observations after action

Usage Examples

Custom Sensor

import numpy as np
from gym import spaces
from habitat.core.simulator import Sensor, SensorTypes

class MyProximitySensor(Sensor):
    cls_uuid = "my_proximity"

    def _get_uuid(self, *args, **kwargs):
        return self.cls_uuid

    def _get_sensor_type(self, *args, **kwargs):
        return SensorTypes.MEASUREMENT

    def _get_observation_space(self, *args, **kwargs):
        return spaces.Box(low=0.0, high=10.0, shape=(1,), dtype=np.float32)

    def get_observation(self, observations, episode, *args, **kwargs):
        agent_pos = self._sim.get_agent_state().position
        goal_pos = np.array(episode.goals[0].position)
        distance = np.linalg.norm(agent_pos - goal_pos)
        return np.array([distance], dtype=np.float32)

Custom Measure

from habitat.core.embodied_task import Measure

class MySuccessMeasure(Measure):
    cls_uuid = "my_success"

    def _get_uuid(self, *args, **kwargs):
        return self.cls_uuid

    def reset_metric(self, episode, task, *args, **kwargs):
        self._metric = 0.0
        self.update_metric(episode=episode, task=task)

    def update_metric(self, episode, task, *args, **kwargs):
        distance = task.measurements.measures["distance_to_goal"].get_metric()
        self._metric = float(distance < self.config.success_distance)

Custom Action

from habitat.core.embodied_task import SimulatorTaskAction

class MyStrafeAction(SimulatorTaskAction):
    def step(self, *args, **kwargs):
        # Move agent sideways
        current_state = self._sim.get_agent_state()
        # ... compute strafe movement ...
        return self._sim.step("strafe_left")

Related Pages

Implements Principle

• Principle:Facebookresearch_Habitat_lab_Component_Class_Definition