Implementation:Facebookresearch Habitat lab HumanoidPickPolicy

Knowledge Sources	Facebookresearch_Habitat_lab
Domains	Embodied_AI, Hierarchical_Reinforcement_Learning
Last Updated	2026-02-15 00:00 GMT

Overview

HumanoidPickPolicy is a hierarchical RL skill policy that generates humanoid picking motions by moving the arm to an object, snapping the hand to it, and retracting, with a companion HumanoidPlacePolicy for placing objects.

Description

HumanoidPickPolicy extends NnSkillPolicy and implements the logic for a humanoid agent to pick up objects. It resolves the target object from PDDL skill arguments via _parse_skill_arg, maps the target entity to an index in the ordered entity list, and produces a humanoid_pick_action with the corresponding entity index and a grab flag. Skill completion is determined by the HasFinishedHumanoidPickSensor. The from_config class method constructs the policy by filtering the action space to only the relevant pick action. The companion HumanoidPlacePolicy inherits from HumanoidPickPolicy and overrides _parse_skill_arg to use RELEASE_ID instead of GRAB_ID, enabling object placement with the same motion infrastructure.

Usage

Use HumanoidPickPolicy as a skill within a hierarchical task planner (HRL) for humanoid rearrangement tasks. It is typically managed by a higher-level policy that selects which skill to execute based on a PDDL plan.

Code Reference

Source Location

• Repository: Facebookresearch_Habitat_lab
• File: habitat-baselines/habitat_baselines/rl/hrl/skills/humanoid_pick.py
• Lines: 21-206

Signature

class HumanoidPickPolicy(NnSkillPolicy):
    GRAB_ID = 1
    RELEASE_ID = 0

    def __init__(
        self,
        wrap_policy,
        config,
        action_space,
        filtered_obs_space,
        filtered_action_space,
        batch_size,
        pddl_domain_path,
        pddl_task_path,
        task_config,
    ):

class HumanoidPlacePolicy(HumanoidPickPolicy):
    def _parse_skill_arg(self, skill_arg):

Import

from habitat_baselines.rl.hrl.skills.humanoid_pick import HumanoidPickPolicy, HumanoidPlacePolicy

I/O Contract

Inputs

Name	Type	Required	Description
wrap_policy	Policy or None	Yes	Wrapped neural network policy (can be None for oracle skills)
config	DictConfig	Yes	Skill configuration containing action_name, skill_name, etc.
action_space	ActionSpace	Yes	Full action space of the environment
filtered_obs_space	Space	Yes	Observation space filtered for this skill
filtered_action_space	ActionSpace	Yes	Action space filtered to the humanoid pick action
batch_size	int	Yes	Number of parallel environments
pddl_domain_path	str	Yes	Path to the PDDL domain definition
pddl_task_path	str	Yes	Path to the PDDL task specification
task_config	DictConfig	Yes	Habitat task configuration

Outputs

Name	Type	Description
action_data	PolicyActionData	Contains the full action tensor with the pick action index and grab/release flag set, plus RNN hidden states

Key Methods

from_config

@classmethod
def from_config(
    cls, config, observation_space, action_space, batch_size, full_config
):

Factory method that constructs the policy from configuration, filtering the action space to the configured action name.

_is_skill_done

def _is_skill_done(
    self, observations, rnn_hidden_states, prev_actions, masks, batch_idx,
) -> torch.BoolTensor:

Returns True when the HasFinishedHumanoidPickSensor indicates the pick action is complete.

_parse_skill_arg

def _parse_skill_arg(self, skill_arg) -> HumanoidPickActionArgs:

Parses PDDL skill arguments to determine the target entity index and grab/release action.

Usage Examples

Basic Usage

from habitat_baselines.rl.hrl.skills.humanoid_pick import HumanoidPickPolicy

# Typically constructed via from_config within HRL framework
pick_skill = HumanoidPickPolicy.from_config(
    config=skill_config,
    observation_space=obs_space,
    action_space=action_space,
    batch_size=num_envs,
    full_config=full_config,
)

# Set the PDDL problem for entity resolution
pick_skill.set_pddl_problem(pddl_problem)

# Enter the skill with PDDL arguments
pick_skill.on_enter(
    skill_arg=["target_obj", "agent_0"],
    batch_idx=[0],
    observations=observations,
    rnn_hidden_states=rnn_states,
    prev_actions=prev_actions,
    skill_name="pick",
)