Implementation:Facebookresearch Habitat lab ResetArmSkill

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

Overview

ResetArmSkill is a hierarchical RL skill that moves a robot arm to a predefined rest joint configuration by computing proportional delta actions normalized by the initial displacement.

Description

ResetArmSkill extends SkillPolicy and implements a simple proportional controller to return the robot arm to a target rest state. Upon entering the skill (on_enter), it records the initial delta between the rest state and the current joint positions. The _internal_act method computes the current delta from the rest state, divides it by the maximum initial delta (clamped to avoid division by zero), and writes the result into the arm action slice of the full action tensor. This normalization ensures actions remain in [-1, 1] and naturally reduces in magnitude as the arm approaches the target. The skill signals completion (_is_skill_done) when the maximum absolute joint error falls below 0.05 radians. The rest joint state is read from config.reset_joint_state.

Usage

Use ResetArmSkill within a hierarchical RL framework to reset the robot arm to its rest position between manipulation subtasks. It is typically invoked by a high-level task planner after a pick or place action.

Code Reference

Source Location

• Repository: Facebookresearch_Habitat_lab
• File: habitat-baselines/habitat_baselines/rl/hrl/skills/reset.py
• Lines: 16-104

Signature

class ResetArmSkill(SkillPolicy):
    def __init__(
        self,
        config,
        action_space: spaces.Space,
        batch_size,
    ):

Import

from habitat_baselines.rl.hrl.skills.reset import ResetArmSkill

I/O Contract

Inputs

Name	Type	Required	Description
config	DictConfig	Yes	Skill configuration containing reset_joint_state (list of target joint positions)
action_space	gym.spaces.Space	Yes	Full action space of the environment, must contain "arm_action"
batch_size	int	Yes	Number of parallel environments

Outputs

Name	Type	Description
action_data	PolicyActionData	Full action tensor with arm delta actions set to drive toward rest state, plus RNN hidden states

Key Methods

on_enter

def on_enter(
    self,
    skill_arg: List[str],
    batch_idxs: List[int],
    observations,
    rnn_hidden_states,
    prev_actions,
    skill_name,
) -> Tuple[torch.Tensor, torch.Tensor]:

Records the initial joint displacement from the rest state for normalization.

_is_skill_done

def _is_skill_done(
    self, observations, rnn_hidden_states, prev_actions, masks, batch_idx
):

Returns True when the maximum joint error is below 0.05 radians.

_internal_act

def _internal_act(
    self, observations, rnn_hidden_states, prev_actions, masks, cur_batch_idx,
    deterministic=False,
):

Computes proportional arm actions normalized by the initial delta.

Usage Examples

Basic Usage

import gym.spaces as spaces
from habitat_baselines.rl.hrl.skills.reset import ResetArmSkill

# Construct the skill from config and action space
reset_skill = ResetArmSkill(
    config=skill_config,        # Must have reset_joint_state
    action_space=env_action_space,  # Must contain "arm_action"
    batch_size=num_envs,
)

# Enter the skill
rnn_states, prev_actions = reset_skill.on_enter(
    skill_arg=[],
    batch_idxs=[0],
    observations=observations,
    rnn_hidden_states=rnn_hidden_states,
    prev_actions=prev_actions,
    skill_name="reset_arm",
)

# Execute the skill in a loop
while not done:
    action_data = reset_skill.act(
        observations, rnn_hidden_states, prev_actions, masks, cur_batch_idx
    )
    observations, rewards, dones, infos = envs.step(action_data.actions)