Implementation:Facebookresearch Habitat lab HumanoidSeqPoseController

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

Overview

HumanoidSeqPoseController is a controller that replays a sequence of pre-recorded humanoid poses loaded from a file, extending HumanoidBaseController with sequential motion playback capabilities.

Description

HumanoidSeqPoseController extends HumanoidBaseController to support loading and sequentially replaying humanoid motion data from a NumPy archive file. On initialization, it loads a motion file containing joint arrays, transform arrays, displacement data, and FPS information, constructing a Motion object from this data.

Key features include:

• Sequential playback -- Supports forward (next_pose) and backward (prev_pose) traversal through the motion sequence, with optional cycling at the boundaries.
• Reference-based positioning -- The first pose in the motion file is set at the provided base transformation, and all subsequent poses are computed relative to that reference.
• Base transformation application -- The apply_base_transformation method allows repositioning the motion sequence by setting a new base offset that makes subsequent poses relative to the given transformation.
• Step size control -- The step size is calculated as the ratio of the motion file's native FPS to the desired playback FPS, allowing temporal resampling of the motion.

The calculate_pose method computes the current joint transforms based on the current motion frame, adjusting the root transform relative to the first pose and the reference pose.

Usage

Use HumanoidSeqPoseController when you need to replay pre-recorded humanoid motion sequences in Habitat simulations. Instantiate it with a path to a .npz motion file, reset it with a base transformation, and then call calculate_pose(advance_pose=True) each step to progress through the animation.

Code Reference

Source Location

• Repository: Facebookresearch_Habitat_lab
• File: habitat-lab/habitat/articulated_agent_controllers/humanoid_seq_pose_controller.py
• Lines: 1-127

Signature

class HumanoidSeqPoseController(HumanoidBaseController):
    def __init__(
        self,
        motion_pose_path,
        motion_fps=30,
        base_offset=(0, 0.9, 0),
    ):

Import

from habitat.articulated_agent_controllers.humanoid_seq_pose_controller import (
    HumanoidSeqPoseController,
)

I/O Contract

Inputs

Name	Type	Required	Description
motion_pose_path	str	Yes	File path to the .npz file containing motion pose data
motion_fps	int	No (default=30)	The FPS at which the pose should be advanced during playback
base_offset	tuple	No (default=(0, 0.9, 0))	Offset between the root of the character and their feet

Outputs

Name	Type	Description
get_pose()	List[float]	Inherited vectorized pose from HumanoidBaseController containing joint poses and transforms
humanoid_motion	Motion	The loaded motion sequence object
motion_frame	int	Current frame index in the motion sequence

Key Methods

Method	Description
reset(base_transformation)	Resets the controller to frame 0 and sets the reference pose to the given base transformation
apply_base_transformation(base_transformation)	Sets a base offset so subsequent poses are relative to the given transformation
next_pose(cycle=False)	Advances to the next pose; optionally cycles back to the beginning at the end
prev_pose(cycle=False)	Moves to the previous pose; optionally cycles to the end when reaching the start
calculate_pose(advance_pose=False)	Computes joint transforms for the current frame; optionally advances to the next pose

Usage Examples

Basic Usage

import magnum as mn
from habitat.articulated_agent_controllers.humanoid_seq_pose_controller import (
    HumanoidSeqPoseController,
)

# Load a motion file and create the controller
controller = HumanoidSeqPoseController(
    motion_pose_path="data/humanoid_motions/walk_motion.npz",
    motion_fps=30,
)

# Reset at a starting transformation
base_transform = mn.Matrix4.identity_init()
controller.reset(base_transform)

# Step through the motion sequence
for step in range(100):
    controller.calculate_pose(advance_pose=True)
    pose = controller.get_pose()
    # Use pose to update humanoid agent in simulation

Related Pages

• Facebookresearch_Habitat_lab_HumanoidBaseController