Load-Aware Quadruped Navigation with Online Contact-based Geometry Estimation

Qi Yin; XiaoLi Zhang; Yangming Shi

cs.ROOct 10, 2026

Load-Aware Quadruped Navigation with Online Contact-based Geometry Estimation

Authors: Qi Yin, XiaoLi Zhang, Yangming Shi

Organizations: Colorado School of Mines, Golden, CO 80401, USA

Abstract

Navigating cluttered environments while carrying objects remains a challenging problem for quadruped robots. Compared with works that focused on load-aware locomotion adaptation to maintain gait stability, existing autonomous navigation systems generally ignore the influence of carried objects' geometries and use fixed footprint inflation, leading to overly conservative navigation. To fill this knowledge gap, we propose a two-phase load-aware navigation learning framework. In Phase 1, we train LOADNav, a load-aware navigation policy via reinforcement learning, enabling the robot to adapt its trajectory and clearance behavior according to the geometric size of the carried object. In Phase 2, we train CoGE, an online geometry predictor via imitation learning, which estimates unknown payload geometry from occasional contact events. We evaluate the framework in cluttered environments with loads of varying sizes and shapes. The results show that our navigation policy exhibits clear load-aware behaviors, and the online geometry predictor estimates object geometry from occasional contact events, refining its estimate and adapting the navigation behavior accordingly. Without requiring any prior knowledge of the payload, it narrows the gap to geometry-aware policies.

Figures & tables

Figure 1: The illustration of showing our proposed two-phase load-aware navigation learning framework. (a) LOADNav (Phase 1): with privileged object size information, the policy learns load-aware navigation strategies, adapting its behavior across sparse and narrow spaces according to different object geometries ( $X<Y$ , $X=Y$ , $X>Y$ ); (b) CoGE-LOADNav (Phase 2): the robot infers object geometry from occasional contact events, enabling adaptive navigation without size inputs.

Figure 2: Overview of our proposed navigation framework. (Left) Three object geometry types categorized by their X and Y dimensions. (Middle) Two-phase training pipeline: Phase 1 trains the load-aware navigation policy with privileged object size input via Reinforcement Learning, while Phase 2 trains the contact-based predictor via Imitation Learning with the main policy network frozen. (Right) Evaluation across simulation, benchmark, and real-world environments.

Figure 3: LOADNav trajectories and velocity profiles for different carried objects (N=80). (a) Tube_X (Length = 1.6 m). (b) Circle (Radius = 0.5 m). (c) Board_Y (Length = 1.6 m).

Figure 4: Visualization of LOADNav and CoGE-LOADNav in simulation, illustrating how CoGE-LOADNav infers geometry through collisions. (a-c) Navigation trajectories of LOADNav and CoGE-LOADNav. (d) Commanded and actual velocities of CoGE-LOADNav. (e-f) Predicted object size and absolute errors.

Figure 5: Memory-enhanced geometry estimation with CoGE-LOADNav+. (a) Trajectories of CoGE-LOADNav and CoGE-LOADNav+. (b) Absolute estimation error of the CoGE-LOADNav+.

Method	Average		Board_Y		Circle		Tube_X
	SR $\uparrow$	ACT $\downarrow$	SR $\uparrow$	ACT $\downarrow$	SR $\uparrow$	ACT $\downarrow$	SR $\uparrow$	ACT $\downarrow$
NavRL	30.00	0.81	30.00	0.33	50.00	0.60	10.00	1.50
VP-Nav	41.67	1.20	20.00	1.50	55.00	0.91	50.00	1.20
VP-Nav-I	28.33	0.46	20.00	0.50	45.00	0.89	20.00	0.00
CoGE-LOADNav	86.67	0.39	80.00	0.50	90.00	0.56	90.00	0.11
CoGE-LOADNav+	88.33	0.60	75.00	0.93	90.00	0.56	100.00	0.30

Table 1: Navigation Performance in the Benchmark.

Method	Average		Board_Y		Tube_X		Tube_Y
	SR $\uparrow$	ACT $\downarrow$	SR $\uparrow$	ACT $\downarrow$	SR $\uparrow$	ACT $\downarrow$	SR $\uparrow$	ACT $\downarrow$
VP-Nav	45.00	1.23	30.00	2.00	50.00	0.60	55.00	1.09
CoGE-LOADNav	81.67	0.33	80.00	0.25	80.00	0.38	85.00	0.35
CoGE-LOADNav+	90.00	0.39	85.00	0.35	95.00	0.37	90.00	0.44
LOADNav	98.33	0.04	95.00	0.11	100.00	0.00	100.00	0.00

Table 2: Navigation Performance in Real-World.

Figure 6: Real-world examples of LOADNav and CoGE-LOADNav.

Figure 7: The results of ablation studies comparing LOADNav, CoGE-LOADNav, w/o NavC., w/o $\mathcal{L}_{obj}$ , and w/o $\mathcal{L}_{a}$ , in terms of success rate, average collision times, and average time.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Reward Term	Weight
Termination	Success reward $r_{\text{succ}}$	400.0
	Timeout penalty $r_{\text{timeout}}$	-40.0
	Robot body contact penalty $r_{\text{robot}}$	-10.0
	Stuck penalty $r_{\text{stuck}}$	-50.0
Dense	Velocity projection reward $r_{\text{vel}}$	6.0
	Heading alignment reward $r_{\text{head}}$	3.0

Appendix

Table 3: Reward Weights.

Parameter	Value
PPO clip parameter	0.2
GAE- $\lambda$	0.95
Reward discount factor	0.99
Learning rate	$3\times 10^{-4}$
Learning rate schedule	adaptive
Value loss coefficient	1.0

Appendix

Table 4: PPO Hyperparameters.

Figure 8: Training environments. (a) Phase 1 with navigation curriculum (N = 20 to 100). (b) Phase 2 without navigation curriculum (N = 80).

Category	Parameter	Value
Terrain	Size	$16\times 16$
Box obstacles	Length range	$0.25\sim 0.55$
	Width range	$0.25\sim 0.55$
	Height range	$0.4\sim 1.0$
Cylinder obstacles	Radius range	$0.15\sim 0.35$
Cylinder obstacles	Height range	$0.4\sim 1.0$

Appendix

Table 5: Navigation Curriculum Parameters.

Object	Parameter	Value
Board_Y	Length range	$0.40\sim 1.60$
	Width range	$0.04\sim 0.20$
	Thickness	0.30
Tube_Y	Length range	$0.40\sim 1.60$
Tube_Y	Radius range	$0.01\sim 0.08$
Circle	Radius range	$0.20\sim 0.50$

Appendix

Table 6: Object Parameters.

Figure 9: Simulation navigation trajectories of LOADNav and CoGE-LOADNav for Tube_X and Board_Y under varying obstacle densities (N = 40, 60, 80, 100).

Figure 10: Warehouse benchmark environment. (a) Simulated warehouse with shelves and 100 randomly placed box obstacles, in which the Unitree Go2 is controlled through ROS2. (b) Occupancy map built online by RTAB-Map and a representative trajectory of LOADNav carrying a Circle payload. (c) The three payloads used in Table 1: Board_Y (X = 0.10 m, Y = 1.60 m), Circle (X = 1.00 m, Y = 1.00 m), and Tube_X (X = 1.60 m, Y = 0.02 m).

Figure 11: Real-world navigation examples of LOADNav and CoGE-LOADNav for Tube_X and Board_Y.

Figure 12: Phase 1 training curves of LOADNav and the LOADNav-B (size-blind policy) (episodes 1500–3000). Left: mean episode reward. Right: mean curriculum level, where a higher level corresponds to a denser obstacle field. Both policies share the curriculum, payload distribution, and PPO settings and differ only in the geometry input.

Robot Navigation

Robot navigation studies how robots perceive, localize, plan, and move safely toward goals in static or changing environments.

See all 532 papers

Feeling Through the Load: Compliant Quadruped Locomotion under Payload Interactions
Oct 6, 2026Shaunak A. Mehta, Mayank Mishra, Prajit KrisshnaKumar +2Robotic Control Compliant Robot Control
Fujitsu Research of America, Pittsburgh, PA, 15213.
TRAIN: Terrain-Aware Interactive Navigation for Quadruped Robots via Reinforcement Learning
Feb 13, 2026Wei Zhu, Irfan Tito Kurniawan, Ye Zhao +1Quadruped Robot Navigation Terrain-Aware Robot Locomotion
Department of Robotics, Graduate School of Engineering, Tohoku University, 980-8579, Sendai, Japan. · Laboratory for Intelligent Decision and Autonomous Robots, Woodruff School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA 30313, USA.
Locomotion with Online Awareness of Dynamics for Precision Footholds
Oct 13, 2026Kunzhao Ren, Zhishuo Zhu, Xiaobin XiongRobot Locomotion RL for Legged Locomotion
Shanghai Innovation Institute. · University of Wisconsin-Madison.

Oct 6, 2026J/K move · Enter open · S save

Feeling Through the Load: Compliant Quadruped Locomotion under Payload Interactions

Shaunak A. Mehta, Mayank Mishra, Prajit KrisshnaKumar, Sebastian Scherer+1

Fujitsu Research of America, Pittsburgh, PA, 15213.

Robotic Control Compliant Robot Control Robot Locomotion Human-Robot Interaction Quadruped Locomotion

Open paper AskPDF arXiv

Quadruped robots are increasingly expected to carry objects while moving through human environments. But what happens when a person interacts directly with the payload rather than with the robot? If the payload is unrestrained, the robot must distinguish intentional external interactions from ordinary payload motion, while still keeping the load balanced and maintaining stable locomotion. How can a quadruped infer and compliantly respond to such interactions using only onboard measurements? In this work, we develop a force-aware locomotion framework that treats payload interactions as commands that shape the motion of the combined robot-payload system. Our approach separates the learning of force-aware locomotion and force estimation on an unrestrained payload. We combine a compliant load-carrying policy with a causal force estimator, trained through estimator-in-the-loop data aggregation and finetuning, to predict interactions from onboard robot measurements. Our simulations and real-world experiments show that the resulting controller can maintain stable payload-carrying locomotion, yield compliantly to external interactions, and use the inferred force to support human-guided changes in the robot's trajectory.

Abstract

Robot Navigation

Explore similar work

Feeling Through the Load: Compliant Quadruped Locomotion under Payload Interactions

TRAIN: Terrain-Aware Interactive Navigation for Quadruped Robots via Reinforcement Learning

Locomotion with Online Awareness of Dynamics for Precision Footholds