Implementation:Facebookresearch Habitat lab Agent ABC

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

Overview

Concrete abstract base class defining the agent interface for Habitat environments, with concrete implementations including PPOAgent and simple baseline agents.

Description

The Agent ABC in habitat-lab defines the `reset()` and `act(observations)` methods. Concrete implementations include:

• PPOAgent: Loads a trained PPO checkpoint and runs inference
• RandomAgent: Takes random actions
• GoalFollower: Uses ShortestPathFollower oracle to navigate to goals

Usage

Subclass `habitat.Agent` to create custom agents, or use `PPOAgent` to wrap trained checkpoints for evaluation.

Code Reference

Source Location

• Repository: habitat-lab
• File: habitat-lab/habitat/core/agent.py (L16-36), habitat-baselines/habitat_baselines/agents/ppo_agents.py (L42-141), habitat-baselines/habitat_baselines/agents/simple_agents.py (L21-44, L75-110)

Signature

class Agent(metaclass=abc.ABCMeta):
    """Abstract base class for agents."""

    @abc.abstractmethod
    def reset(self) -> None:
        """Called at the beginning of each episode."""
        raise NotImplementedError

    @abc.abstractmethod
    def act(self, observations: Observations) -> Union[int, str, Dict[str, Any]]:
        """
        Args:
            observations: Current observations from environment
        Returns:
            Action to execute
        """
        raise NotImplementedError

# PPOAgent implementation
class PPOAgent(Agent):
    def __init__(self, config: DictConfig):
        """Load trained PPO policy from checkpoint."""

    def reset(self) -> None:
        """Reset recurrent hidden states."""

    def act(self, observations) -> Dict[str, int]:
        """Run policy inference."""

Import

from habitat.core.agent import Agent
from habitat_baselines.agents.ppo_agents import PPOAgent
from habitat_baselines.agents.simple_agents import RandomAgent, GoalFollower

I/O Contract

Inputs

Name	Type	Required	Description
observations	Observations (Dict)	Yes	Current sensor observations from environment
config	DictConfig	Yes (for PPOAgent)	Config with checkpoint path

Outputs

Name	Type	Description
action	Union[int, str, Dict]	Action to execute in the environment

Usage Examples

Create PPO Agent

from habitat_baselines.agents.ppo_agents import PPOAgent

agent = PPOAgent(config)
agent.reset()
action = agent.act(observations)

Create Random Agent

from habitat_baselines.agents.simple_agents import RandomAgent

agent = RandomAgent(
    success_distance=0.2,
    goal_sensor_uuid="pointgoal_with_gps_compass",
)

Related Pages

Implements Principle

• Principle:Facebookresearch_Habitat_lab_Agent_Implementation

Requires Environment

• Environment:Facebookresearch_Habitat_lab_Python_3_9_Core_Dependencies