SkateNav: Learning Energy-Efficient Navigation through Adaptive Skating on Passive-Wheel Quadrupedal Robots

Jeil Jeong; Taeyun Kim; Sangjin Kim; Jinyoung Choi; Sung-Eui Yoon

cs.ROOct 19, 2026

SkateNav: Learning Energy-Efficient Navigation through Adaptive Skating on Passive-Wheel Quadrupedal Robots

Authors: Jeil Jeong, Taeyun Kim, Sangjin Kim, Jinyoung Choi, Sung-Eui Yoon

Organizations: NAVER LABS, Republic of Korea. · School of Computing, KAIST, Republic of Korea.

Abstract

Passive-wheel quadrupedal robots offer energy-efficient locomotion through leg-driven propulsion and momentum-driven gliding. However, navigation requires coordinating steering with intermittent propulsion and adapting push-glide timing to task demands. Velocity-tracking objectives may suppress the speed variations inherent in skating, discouraging efficient gliding. To address these challenges, we propose SkateNav, a unified skating-navigation framework based on reinforcement learning. A time-budgeted goal-reaching objective allows the policy to jointly learn steering, propulsion, and their temporal allocation. An adaptive skating cycle structures gliding, pushing, and recovery, while frequency modulation at every control step adjusts phase durations to navigation demands. Simulation experiments demonstrate higher average success rates and lower cost of transport than the evaluated planner-controller baselines, with ablations confirming the contributions of cycle structure and frequency adaptation. Real-world experiments demonstrate zero-shot transfer from simulation to a passive-wheel Mini Cheetah across cluttered indoor spaces, slopes, and paved outdoor surfaces. Supplementary videos are available on the project page: https://ktw1404.github.io/SkateNav/

Figures & tables

Figure 1: SkateNav for passive-wheel navigation in diverse environments. Our SkateNav framework enables local navigation of a Mini Cheetah [ 20 ] equipped with custom detachable passive-wheel modules. Representative snapshots show the robot navigating toward the goal through (a) gliding on passive wheels, (b) obstacle avoidance in a cluttered passage, (c) heading adjustment, and (d, e) uphill and downhill slope traversal.

Figure 2: Mechanical design and hardware implementation. (a) Overall design of the quadruped robot equipped with passive-wheel modules. (b,c) Side and assembly views of the foot and detachable passive-wheel module. (d) Hardware implementation, with the blue and orange boxes indicating foot-ground and wheel-ground contact modes, respectively.

Figure 3: Overview of the proposed SkateNav framework. A unified skating-navigation policy uses a 2.5D elevation map, robot states, and navigation commands to generate joint actions and a frequency action. The navigation commands specify the current and next waypoints and the time remaining to reach the current waypoint. The frequency action controls progression of the skating cycle, allowing phase durations to vary with navigation demands.

Figure 4: Each panel shows the cumulative distributions of instantaneous mechanical power for policies trained with goal-reaching (solid) and velocity-tracking (dashed) objectives on each platform, averaged over 100 trials. Both policies traverse

7\text{\,}\mathrm{m}

7\text{\,}\mathrm{s}

, with a

1\text{\,}\mathrm{m}

command for velocity tracking. The vertical axis shows the percentage of time at or below each power level; shading highlights the difference in the low-power regime.

Figure 5: (a) Each leg sequentially follows Glide, Push, and Recovery in the phase domain, with the right and left leg pairs offset by half a cycle. (b) As the cycle progresses, the desired contact configurations are defined: while the right-side legs remain in Glide, the left-side legs transition through Glide, Push, and Recovery with wheel, foot, and no contact, respectively.

Figure 6: (a) A fixed skating frequency advances the phase with predefined durations, while (b) an adaptive skating frequency is adjusted at every control step, allowing durations to vary according to navigation demands.

Parameter	Train Distribution	Test Distribution	Unit
Payload	$\mathcal{U}(0.0,\,3.0)$	$\mathcal{U}(0.0,\,5.0)$	$\mathrm{kg}$
CoM Offset	$\mathcal{U}(-0.1,\,0.1)^{3}$	$\mathcal{U}(-0.15,\,0.15)^{3}$	$\mathrm{m}$
Wheel Friction	$\mathcal{U}(5\mathrm{e}{-4},\,4\mathrm{e}{-3})$	$\mathcal{U}(0.0,\,1\mathrm{e}{-2})$	$\mathrm{N}\text{\,}\mathrm{m}$
Wheel Armature	$\mathcal{U}(2\mathrm{e}{-5},\,1\mathrm{e}{-4})$	$\mathcal{U}(0.0,\,1\mathrm{e}{-3})$	$\mathrm{kg}\text{\,}{\mathrm{m}}^{2}$
Terrain Friction	$\mathcal{U}(0.7,\,1.2)$	$\mathcal{U}(0.4,\,1.5)$	–
Terrain Restitution	$\mathcal{U}(0.0,\,0.1)$	$\mathcal{U}(0.0,\,0.3)$	–

Table 1: Domain Randomization Parameters

Reward Term	Expression
Task Rewards : $\mathcal{R}_{\text{task}}=\sum_{k=1}^{2}r_{k}$
Waypoint Arrival ( $r_{1}$ )	$w_{1}\;\mathds{1}_{t_{\text{seg}}>T_{\text{seg}}-T_{r}}/[T_{r}(1+\\|\boldsymbol{G}^{\mathcal{B}}_{\mathrm{cur}}\\|_{2}^{2})]$
Skating Gait ( $r_{2}$ )	$w_{2}\;(\sum_{i=0}^{3}\left[S_{W}(\phi_{i})I_{W_{i}}+S_{F}(\phi_{i})I_{F_{i}}\right])/N_{\mathrm{des}}$

Table 2: Skating-Navigation Policy Rewards

\mathcal{R}=\mathcal{R}_{\text{task}}+\mathcal{R}_{\text{reg}}

	Flat		BARN-easy		BARN-hard		Uphill		Downhill		Average		IT
Method	SR $\uparrow$	CoT $\downarrow$	SR $\uparrow$	CoT $\downarrow$	SR $\uparrow$	CoT $\downarrow$	SR $\uparrow$	CoT $\downarrow$	SR $\uparrow$	CoT $\downarrow$	SR $\uparrow$	CoT $\downarrow$	[ms]
Ours	100.0	0.512	0 89.3	0.539	0 79.0	0.554	100.0	0.976	0 98.4	0.216	0 93.3	0.559	0.28
$\hookrightarrow$ w/o Adaptive Freq.	100.0	0.628	0 88.0	0.695	0 78.0	0.653	100.0	1.402	0 70.0	0.348	0 87.2	0.745	0.28
$\hookrightarrow$ w/o Skating Cycle	0 83.3	1.019	0 75.5	1.058	0 72.8	1.067	0 96.3	1.689	0 84.8	0.428	0 82.5	1.052	0.27
DWA+VT [ 11 ]	100.0	0.608	0 73.1	0.657	0 55.5	0.683	0 99.8	1.046	0 97.7	0.341	0 85.2	0.667	0.42
MPPI+VT [ 36 ]	0 93.4	0.606	0 77.4	0.650	0 70.9	0.664	0 86.9	1.093	0 94.1	0.282	0 84.5	0.659	1.31

Table 3: Navigation performance in simulation environments. For details of experimental setups, refer to Sec. 6.1 and Sec. 6.2 . For w/o Adaptive Freq., we report results at the fixed frequencies yielding the highest SR:

2.0

\mathrm{Hz}

( Flat / BARN-easy ),

1.8

\mathrm{Hz}

( BARN-hard / Uphill ), and

1.2

\mathrm{Hz}

( Downhill ). For planners, we select the best cost setting (aggressive, balanced, or safe) by average SR: balanced for DWA; aggressive for MPPI and CEM.

Figure 7: (a) Flat . (b)–(c) BARN-easy and BARN-hard , with obstacle layouts based on BARN [ 28 ] configurations and mean clearances of $4.26$ and $2.16$ $\mathrm{m}$ , respectively. (d) The $15^{\circ}$ slope for Uphill and Downhill conditions.

Figure 8: Mean skating frequency collected (a) during the first segment of each navigation task across waypoint distances and terrain conditions, and (b) grouped by observed obstacle clearance from the robot center.

Figure 9: Each task comprises four

5

\mathrm{m}

3.5

\mathrm{s}

segments, with MPPI+VT commanding

\approx 1.4

\mathrm{m}

on average. (a) Phase durations and skating frequency over normalized cycle time. R/L denote right/left leg pairs;

\uparrow

\downarrow

indicate longer/shorter VT phase durations than Ours. (b) Energy proxy and traveled distance accumulated within each phase pair over the evaluation.

Figure 10: Frequency sweep across representative conditions. Green boxes show CoT distributions for w/o Adaptive Freq. at selected fixed frequencies; lines show medians across the sweep. Magenta boxes show Ours at its mean skating frequency. Numbers above boxes indicate SR (%).

Figure 11: Real-world navigation example using a passive-wheel quadruped robot. The point-cloud map in the middle row shows the robot trajectory colored by body velocity. Markers A–F indicate selected locations along the route, which includes curved passages, a downhill slope, and cluttered indoor spaces. The yellow inset provides a top-down view of the marked section. The top row shows the robot at each location, and the bottom row shows the corresponding local elevation map provided to the policy, with green and yellow spheres indicating the current and next waypoints, respectively.

Figure 12: (a) Hardware setup . (b) Real-world evaluation environments .

Robot Locomotion

Robot locomotion studies how robots move efficiently and reliably across varied terrains, environments, and movement modes.

See all 225 papers

Efficient and Versatile Quadrupedal Skating: Optimal Co-design via Reinforcement Learning and Bayesian Optimization
Mar 19, 2026Hanwen Wang, Zhenlong Fang, Josiah Hanna +1Robot Morphology-Control Co-Design Quadruped Locomotion
Hanwen Wang, Zhenlong Fang, and Josiah Hanna are with the University of Wisconsin - Madison. · Xiaobin Xiong was with the University of Wisconsin - Madison, and now with the Shanghai Innovation Institute (SII).
Reinforcement Learning-Based Control for an Inline Skating Humanoid Robot
Jun 30, 2026Ethan Marot, Thomas Bi, Clemens Schwarke +3Humanoid Robot Locomotion Robot Locomotion
Institute for Dynamic Systems and Control, ETH Zurich, Switzerland. · Robotic Systems Lab, ETH Zurich, Switzerland
SKATER: Synthesized Kinematics for Advanced Traversing Efficiency on a Humanoid Robot via Roller Skate Swizzles
Jan 8, 2026Junchi Gu, Feiyang Yuan, Weize Shi +6Robot Locomotion Humanoid Robot Locomotion
Institute of Humanoid Robots, Department of Precision Machinery and Precision Instrumentation, University of Science and Technology of China, Hefei, Anhui 230026, China.

Mar 19, 2026J/K move · Enter open · S save

Efficient and Versatile Quadrupedal Skating: Optimal Co-design via Reinforcement Learning and Bayesian Optimization

Hanwen Wang, Zhenlong Fang, Josiah Hanna, Xiaobin Xiong

Hanwen Wang, Zhenlong Fang, and Josiah Hanna are with the University of Wisconsin - Madison. · Xiaobin Xiong was with the University of Wisconsin - Madison, and now with the Shanghai Innovation Institute (SII).

Robot Morphology-Control Co-Design Quadruped Locomotion Energy-Efficient Robot Locomotion

Open paper AskPDF arXiv

In this paper, we present a hardware-control co-design approach that enables efficient and versatile roller skating on quadrupedal robots equipped with passive wheels. Passive-wheel skating reduces leg inertia and improves energy efficiency, particularly at high speeds. However, the absence of direct wheel actuation tightly couples mechanical design and control. To unlock the full potential of this modality, we formulate a bilevel optimization framework: an upper-level Bayesian Optimization searches the mechanical design space, while a lower-level Reinforcement Learning trains a motor control policy for each candidate design. The resulting design-policy pairs not only outperform human-engineered baselines, but also exhibit versatile behaviors such as hockey stop (rapid braking by turning sideways to maximize friction) and self-aligning motion (automatic reorientation to improve energy efficiency in the direction of travel), offering the first system-level study of dynamic skating motion on quadrupedal robots.

Abstract

Robot Locomotion

Explore similar work

Efficient and Versatile Quadrupedal Skating: Optimal Co-design via Reinforcement Learning and Bayesian Optimization

Reinforcement Learning-Based Control for an Inline Skating Humanoid Robot

SKATER: Synthesized Kinematics for Advanced Traversing Efficiency on a Humanoid Robot via Roller Skate Swizzles