Camera-to-Quadruped Robot Pose Estimation from Sequential Images

Seobin Hwang; Ji-Seong Beom; Yeong-Jun Cho

cs.ROOct 27, 2026

Camera-to-Quadruped Robot Pose Estimation from Sequential Images

Authors: Seobin Hwang, Ji-Seong Beom, Yeong-Jun Cho

Organizations: Department of Artificial Intelligence Convergence, Chonnam National University, Gwangju, Republic of Korea

Abstract

Quadruped robots are increasingly deployed in hazardous environments such as disaster response and facility inspection, where external-camera monitoring is essential for safe operation. This setting is challenging because quadrupeds are articulated systems with time-varying leg joints, leg symmetry, and frequent self-occlusions, which make keypoint correspondence ambiguous. Moreover, no public benchmark exists, making quantitative comparison difficult. To address these challenges, we propose QuadPose, which estimates the full-body 18-DoF pose (6-DoF base pose + 12-DoF joint angles) from monocular video sequences, and introduce QuadSyn, the first camera-to-quadruped robot pose benchmark. On QuadSyn, QuadPose achieves 92.18% [email protected], 95.54% PCK-AUC, 85.94% ADD@20cm, 66.98% ADD-AUC, and 3.65 degree mean joint-angle error at approximately 33 FPS.

Figures & tables

Figure 1: Overview of camera-to-quadruped robot pose estimation. (a) Problem setup and challenges: an articulated quadruped model (6-DoF base + 12-DoF joints) and 2D keypoint ambiguities caused by self-occlusion and leg symmetry. (b) Overview of the proposed QuadPose framework and the QuadSyn dataset.

Category	Method	Robot type	Input	Output *
	RoboPose (CVPR’21) [ 12 ]	Manipulator	Image	6DoF + $q(7)$
Mesh	DronePose (ECCVW’20) [ 1 ]	Drone	Image	6DoF
	DroneKey++ (ICRA’26) [ 9 ]	Drone	Image seq.	2D kpt + 6DoF
	RoboPEPP (CVPR’25) [ 6 ]	Manipulator	Image	2D/3D kpt + 6DoF + $q(7)$
	DroneKey (IROS’25) [ 8 ]	Drone	Image seq.	2D kpt + 6DoF
	SGTAPose (CVPR’23) [ 22 ]	Manipulator	Image seq.	2D/3D kpt + 6DoF

Table 1: Overview of camera-to-robot pose estimation methods. Existing methods are categorized into mesh-based and keypoint-based approaches, and we summarize their target robot types, input modalities, and outputs.

Figure 2: The overall framework of QuadPose . (a) 2D keypoint detector with selector tokens and skeleton-aware masked attention. (b) 3D pose estimator that predicts joint angles and base pose, with FK reconstruction and refinement. (c) Loss functions.

Figure 5: QuadSyn dataset annotations and examples. (a) 2D bounding box and keypoint labels. (b) 3D keypoints, joint angles, and 6-DoF base pose in the camera frame. (c) Scene examples with overlaid leg keypoint trajectories (30 frames) and robot paths.


	PCK @ (pix)					ADD @ (cm)
Method	2.5 $\uparrow$	5 $\uparrow$	10 $\uparrow$	AUC $\uparrow$	Med $\downarrow$	10 $\uparrow$	15 $\uparrow$	20 $\uparrow$	AUC $\uparrow$	Med $\downarrow$

DREAM (ICRA’20) [ 13 ]	24.45	75.62	98.52	61.13	3.720	19.10	28.20	35.70	8.400	31.93
RoboPose (CVPR’21) [ 12 ]	–	–	–	–	–	0.120	0.450	0.720	0.004	732.6
YOLOv8-pose (’23) [ 10 ]	70.44	92.54	98.97	88.90	1.683	–	–	–	–	–

Table 2: PCK and ADD comparison on the QuadSyn benchmark. We compare with state-of-the-art methods [ 13 , 12 , 10 , 22 , 6 ] and report PCK at 2.5/5/10 px and ADD at 10/15/20 cm, each with AUC and median error (

\uparrow

: higher is better,

\downarrow

: lower is better). Missing entries (–) indicate that the method does not produce the corresponding output.


	Joint Error @ (degree) $\downarrow$
	Front-Left			Front-Right			Rear-Left			Rear-Right
Method	J1	J2	J3	J4	J5	J6	J7	J8	J9	J10	J11	J12

RoboPose (CVPR’21) [ 12 ]	10.6	10.6	9.56	9.45	19.9	14.5	13.1	19.2	12.7	25.0	13.2	8.98
RoboPEPP (CVPR’25) [ 6 ]	4.11	4.46	4.01	3.85	3.85	3.99	4.14	3.61	4.09	4.40	3.97	4.33

Table 3: Joint-angle error comparison on the QuadSyn benchmark. We compare with RoboPose [ 12 ] and RoboPEPP [ 6 ] and report per-joint angle error (

\downarrow

: lower is better).

Figure 6: Qualitative comparison of 2D keypoint detection on QuadSyn .

Figure 7: Visualization of 3D pose estimation results on QuadSyn . QuadPose (ours) vs. RoboPEPP [ 6 ] . Colors indicate error. MPJPE (Mean Per Joint Position Error) in cm.

Table 10

Robot State Estimation

Robot state estimation infers a robot’s position, motion, orientation, and internal configuration from noisy sensors and physical constraints.

See all 200 papers

Panoramic Multimodal Semantic Occupancy Prediction for Quadruped Robots
Mar 13, 2026Guoqiang Zhao, Zhe Yang, Sheng Wu +53D Occupancy Prediction Robotic Perception
The authors are with the School of Artificial Intelligence and Robotics, Hunan University, Changsha 410012, China · The author is also with the National Engineering Research Center of Robot Visual Perception and Control Technology, Hunan University, Changsha 410082, China.
CopyQuad: Quadrupedal Skill Transfer from a Single Side-View Video Without a Depth Model
Oct 12, 2026Mohamad H. Danesh, Raina Hu, Mengxue Xu +1Robot Imitation Learning Quadruped Locomotion
McGill University · Mila – Quebec AI Institute
QuadVerse: An Integrated Framework Aligning Visual-Physical Reality for Quadruped Simulation
Jun 5, 2026Yuxiang Chen, Yuanhao Wang, Ziheng Zhang +6Robotics Simulation Sim-to-Real Transfer
1Nanjing University · 2BUPT · 3DEXMAL +1

Mar 13, 2026J/K move · Enter open · S save

Panoramic Multimodal Semantic Occupancy Prediction for Quadruped Robots

Guoqiang Zhao, Zhe Yang, Sheng Wu, Fei Teng+4

The authors are with the School of Artificial Intelligence and Robotics, Hunan University, Changsha 410012, China · The author is also with the National Engineering Research Center of Robot Visual Perception and Control Technology, Hunan University, Changsha 410082, China.

3D Occupancy Prediction Robotic Perception Multimodal Sensor Fusion

Open paper AskPDF arXiv

Panoramic imagery provides holistic 360° visual coverage for environmental perception in quadruped robots. However, existing occupancy prediction methods are primarily designed for wheeled autonomous driving and rely heavily on RGB cues, which limits their robustness in complex, dynamically changing environments. To bridge this gap, we introduce PanoMMOcc, the first real-world panoramic multimodal occupancy dataset for quadruped robots, comprising four sensing modalities collected across diverse scenes. We further propose VoxelHound, a panoramic multimodal occupancy perception framework tailored to legged locomotion and spherical imaging. VoxelHound incorporates a Vertical Jitter Compensation (VJC) module to mitigate severe viewpoint perturbations caused by body pitch and roll during locomotion, enabling more consistent spatial reasoning, and a Multimodal Information Prompt Fusion (MIPF) module to effectively integrate panoramic visual cues with auxiliary modalities for enhanced volumetric occupancy prediction. We also establish a comprehensive benchmark on PanoMMOcc and provide detailed dataset analyses to enable systematic evaluation in challenging embodied perception scenarios. Extensive experiments demonstrate that VoxelHound achieves state-of-the-art performance on PanoMMOcc, with a +4.16 gain in mIoU. The dataset and code will be publicly released to facilitate future research on panoramic multimodal 3D perception for embodied robotic systems at https://github.com/SXDR/PanoMMOcc.

Abstract

Robot State Estimation

Explore similar work

Panoramic Multimodal Semantic Occupancy Prediction for Quadruped Robots

CopyQuad: Quadrupedal Skill Transfer from a Single Side-View Video Without a Depth Model

QuadVerse: An Integrated Framework Aligning Visual-Physical Reality for Quadruped Simulation