Implementation:Facebookresearch Habitat lab PointNavResNetPolicy from config

Knowledge Sources	Habitat-Lab DD-PPO
Domains	Deep_Learning, Reinforcement_Learning, Computer_Vision
Last Updated	2026-02-15 02:00 GMT

Overview

Concrete factory method for constructing the ResNet-based visual navigation policy from a Habitat config, provided by the habitat-baselines package.

Description

The PointNavResNetPolicy.from_config class method constructs a complete actor-critic policy network for point-goal navigation. It instantiates a ResNetEncoder for processing visual observations, a RNNStateEncoder (GRU by default) for temporal integration, and linear action/value heads. The architecture is configured via the experiment config, supporting ResNet-18/50 backbones, variable hidden sizes, and optional visual input normalization.

Usage

This method is called during `PPOTrainer._init_train()` to construct the policy network. It reads architecture parameters from config and adapts to the environment's observation and action spaces.

Code Reference

Source Location

• Repository: habitat-lab
• File: habitat-baselines/habitat_baselines/rl/ddppo/policy/resnet_policy.py
• Lines: L115-162 (from_config), L402-603 (PointNavResNetNet.__init__)

Signature

class PointNavResNetPolicy(NetPolicy):
    @classmethod
    def from_config(
        cls,
        config: DictConfig,
        observation_space: spaces.Dict,
        action_space,
        **kwargs,
    ) -> "PointNavResNetPolicy":
        """
        Construct policy from config.

        Args:
            config: DictConfig with policy architecture parameters
            observation_space: Dict space from environment
            action_space: Discrete or continuous action space
        Returns:
            PointNavResNetPolicy instance
        """

Import

from habitat_baselines.rl.ddppo.policy.resnet_policy import PointNavResNetPolicy

I/O Contract

Inputs

Name	Type	Required	Description
config	DictConfig	Yes	Experiment config with `habitat_baselines.rl.policy` section
observation_space	spaces.Dict	Yes	Environment observation space (defines which visual inputs exist)
action_space	Space	Yes	Environment action space (discrete or continuous)

Outputs

Name	Type	Description
policy	PointNavResNetPolicy	Complete actor-critic policy with ResNet encoder + GRU + action/value heads

Usage Examples

Construct Policy from Config

from habitat_baselines.rl.ddppo.policy.resnet_policy import PointNavResNetPolicy

# Construct policy from config (typically done inside PPOTrainer._init_train)
policy = PointNavResNetPolicy.from_config(
    config=config,
    observation_space=envs.observation_spaces[0],
    action_space=envs.action_spaces[0],
)
policy.to(device)

Related Pages

Implements Principle

• Principle:Facebookresearch_Habitat_lab_Policy_Network_Initialization

Requires Environment

• Environment:Facebookresearch_Habitat_lab_CUDA_GPU_Training_Environment