RAIWAY: A Large-Wheeled Bipedal Robot with Contact-Informed Blind Locomotion

Kyungmin Nam; Jiseong Lee; Donghyuk Choi; Jeongsoo Park; Woojin Kim; Jemin Hwangbo

cs.ROOct 23, 2026

RAIWAY: A Large-Wheeled Bipedal Robot with Contact-Informed Blind Locomotion

Authors: Kyungmin Nam, Jiseong Lee, Donghyuk Choi, Jeongsoo Park, Woojin Kim, Jemin Hwangbo

Organizations: Robotics and Artificial Intelligence Lab, Korea Advanced Institute of Science & Technology (KAIST), Daejeon 34141, South Korea

Abstract

We present RAIWAY, an in-house wheeled bipedal robot with 20-inch wheels, together with a contact-informed control approach for rolling locomotion across irregular and discontinuous terrain. Such locomotion is fundamentally contact-dominated, governed by local wheel--terrain interaction that is impractical to instrument directly and depends on more than terrain geometry alone. The proposed framework exploits the contact information available in simulation, distilling multi-contact events into a latent representation. At deployment, the controller infers this representation online from proprioception, requiring no contact sensing. In real-world experiments, the proposed system traverses diverse environments including steep slopes, gravel beds, studded surfaces, and rocky terrain, and demonstrates continuous ascent on the tested outdoor staircases with riser heights up to 15~cm. On our platform, the proposed controller further outperforms baselines without the contact-informed mechanism across challenging indoor and outdoor settings. The video can be found at https://youtu.be/Ch0divLg64w.

Figures & tables

Fig. 1: Real-world experiments across diverse challenging environments: Representative deployments include (A) rapid stair descent, (B,C) steep-slope descent and ascent, traversal of (D) gravel bed, (E) studded slope, and (F,G) rocky terrain, and continuous stair ascent at (H) 9 cm, (I) 14 cm, and (J) 15 cm riser heights.

Fig. 2: Design of RAIWAY: The main view shows the overall geometry and key dimensions, with insets detailing the leg mechanism, the wheel-actuator mount, and a cross-section of the closed-cell foam tire insert.

Fig. 3: Overview of the proposed contact-informed policy learning framework: The large wheels of our platform produce informative multi-contact patterns, yielding a diverse set of spatially distributed contact events during training. These noisy multi-contact events from the simulator are assembled into a structured vector $c_{t}$ . A recurrent encoder maps this vector sequence into a latent representation, which a student module learns to infer from onboard proprioceptive history for deployment. The GRU encoder has a hidden size of $128$ , the estimator and policy are $[256,128]$ MLPs, the teacher encoder is an LSTM with a hidden size of $8$ producing the $32$ -dimensional latent $\mathbf{l}_{t}$ , and the student is a $[256,128]$ MLP, all with LeakyReLU activations.

Fig. 4: Temporal filtering during a representative simulated rollout: The robot traverses a flat–stairs–drop–flat terrain sequence, where the flat sections feature low-amplitude fractal height variations. The graphs compare the $x_{s}$ -coordinate of the first angularly sorted contact position for the two wheels before and after temporal filtering. On the flat terrain, the raw signal flickers despite continuous rolling contact, whereas the filtered signal preserves variation. During stair ascent, the filtered signal captures the step-edge locations with a distinct piecewise evolution. During the drop segment, a true aerial phase is induced, and the debounce filter introduces a short delay in the contact-state transition.

Category	Parameter	Range / Unit
Wheel Collision Body	Radius	$+\,\mathcal{U}(-4,\,0)~[\mathrm{cm}]$
	Thickness	$+\,\mathcal{U}(-4,\,0)~[\mathrm{cm}]$
	Center offset	$+\,\mathcal{U}(-2,\,2)~[\mathrm{cm}]$
	Mass	$+\,\mathcal{U}(-0.05,\,0.20)~[\mathrm{kg}]$
	Inertia	$+\,\mathcal{U}(-0.02,\,0.03)~[\mathrm{kg\!\cdot\!m^{2}}]$
Terrain Parameters	Restitution coefficient	$\mathcal{U}(0.2,\,0.3)$

TABLE I: Contact-Related Domain Randomization Ranges

Reward	Expression	Coeff.
Forward Velocity Tracking	$\exp\!\big(-({v}_{x}^{des}-{v_{x}}^{\prime})^{2}\big),\quad{v_{x}}^{\prime}=\operatorname{sign}(v_{x})\,v_{xz}$	$1.5$
Lateral Velocity Tracking	$\exp\!\big(-0.5{v_{y}}^{2}\big)$	$1.0$
Angular Velocity Tracking	$\exp\!\big(-1.5({\omega}_{z}^{des}-\omega_{z})^{2}\big)$	$1.0$
Orientation	$R_{zz}$	$1.0$
Foot Slip	$\sum_{i\in\{L,R\}}\mathbf{1}[i\text{ in contact}]\,\big\|\,v_{f,xz,i}-r_{i}\,\|\Omega_{i}\|\,\big\|$	$-0.5$
Contact	$\sum_{i\in\{L,R\}}\mathbf{1}[i\text{ not in contact}]$	$-0.1$

TABLE II: Reward Functions

Variant	Traversal Dist. [%]	Variant	Traversal Dist. [%]
Ours	74.4	w/o contact position	50.4
w/o wheel-level features	64.1	2 slots	46.2
AC + estimator	57.2	1 slot	44.2
w/o contact impulse	53.6	w/o contact pos & imp	42.9

TABLE III: Contact-vector ablation in simulation

Fig. 5: t-SNE projection of $\mathbf{l}_{t}$ . The left panel shows the latent at mid-training before the regularization term was activated ( $\lambda=0$ ), and the right panel shows the latent at the end of training ( $\lambda=1$ ).

	Components		Experiment 1			Experiment 2
Controller	Contact- informed	Estimator	Fail	Unstable	Success	Traversal Dist. [%], mean [min, max]
Plain AC	X	X	6	9	5	69 [35, 100]
AC + estimator	X	O	4	2	14	39 [5, 65]
Ours w/o Estimator	O	X	2	4	14	83 [25, 100]
Ours w/o Filtering	O	O	2	1	17	81 [25, 100]
Ours	O	O	0	0	20	95 [75, 100]

TABLE IV: Ablation study results

Fig. 6: Temporal response during a 16 cm step ascent: Velocities and torques of the left leg joints, together with the latent distance $d_{t}$ , aligned to the ascent onset ( $t=0$ ). The last panel shows the representative contact location and wheel rotation, with the dashed horizontal line indicating the geometrically required wheel rotation for the 16 cm step ascent.

Panel of Figure 11 — Fig. 7: Responses to abrupt contact events during outdoor deployment: Two representative episodes involving challenging contact conditions. The plots show base angle, leg-joint positions, the event-relevant estimator output, and two latent components ( $l_{3},\,l_{4}$ ). (a) The left wheel spins against a wheel-radius-scale rock in (i)–(ii), after which the left leg rolls over the rock on its wheel in (iii), producing a pronounced body oscillation after landing in (iv). (b) The studded slope shifts the wheel–terrain contact locations laterally.

Label	Environment	Dimensions	Command [m/s]
A	Stair descent	18 cm (H) $\times$ 30 cm (W), 21 steps	3.0
B	Downhill slope	40 ∘ incline	2.0
C	Uphill slope	20 ∘ incline	2.5
D	Gravel bed	6–8 cm stones	3.0
E	Studded slope	20 ∘ incline, studs of 5 cm height	1.5
F	Rocky terrain	$\leq$ 22 cm (H), 25–50 cm (W)	1.0 †

TABLE V: Summary of real-world environments

Robot Locomotion

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

See all 233 papers

MILD: Tractable Terrain Modeling for Learning Improved Bipedal Locomotion on Deformable Surfaces
Aug 20, 2026Zeren Luo, Jiahui Zhang, Zhe Xu +6Terrain-Aware Robot Locomotion Humanoid Robot Locomotion
Adaptive Robotic Controls Lab (ArcLab), Department of Mechanical Engineering, The University of Hong Kong, Hong Kong · School of Mechanical and Electrical Engineering, Beijing Institute of Technology, Beijing, China · Graduate School of Information Science and Technology, The University of Tokyo, Tokyo, Japan
Multi-Terrain Mastery: A Comprehensive Controller for Bipedal Locomotion
Sep 27, 2026Oluwami Dosunmu-Ogunbi, Aayushi ShrivastavaRobot Locomotion Legged Robotics
Mechanical Engineering Department Ohio Northern University Ada, United States of America · Mechanical Engineering Department University of California, Berkeley Berkeley, United States of America
Robust bipedal locomotion on flowable slopes via foot-driven terrain manipulation
Jul 13, 2026Deniz Kerimoglu, Junnosuke Kamohara, Jiyeon Maeng +4Humanoid Robot Locomotion Granular Terrain Locomotion
Department of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA · School of Physics, Georgia Institute of Technology, Atlanta, GA 30332, USA · Department of Electrical and Computer Engineering, Northeastern University, Boston, MA 02115, USA

Aug 20, 2026J/K move · Enter open · S save

MILD: Tractable Terrain Modeling for Learning Improved Bipedal Locomotion on Deformable Surfaces

Zeren Luo, Jiahui Zhang, Zhe Xu, Wanyue Li+5

Adaptive Robotic Controls Lab (ArcLab), Department of Mechanical Engineering, The University of Hong Kong, Hong Kong · School of Mechanical and Electrical Engineering, Beijing Institute of Technology, Beijing, China · Graduate School of Information Science and Technology, The University of Tokyo, Tokyo, Japan

Terrain-Aware Robot Locomotion Humanoid Robot Locomotion Physics-Based Simulation Robot Policy Adaptation RL for Legged Locomotion

Open paper AskPDF arXiv

Enabling robots to walk on yielding terrain is vital for applications ranging from disaster response to planetary exploration. While bipedal robots hold immense potential, their locomotion on deformable surfaces remains limited as current simulators fail to capture the spatiotemporal heterogeneity of such yielding substrates. We present MILD, featuring a physics-grounded discrete-element contact solver that accurately simulates spatially varying foot-terrain interactions. Complementing this model, we train a terrain-aware locomotion controller via deep reinforcement learning with latent modulation and proprioceptive estimation. Quantitative comparisons against state-of-the-art methods show our approach generates more diverse and realistic contact scenarios during training, resulting in controllers that exhibit natural adaptation on real deformable surfaces. Through hardware experiments, we demonstrate the system's capability for online terrain identification and adaptation across a wide range of surface stiffness.

Abstract

Robot Locomotion

Explore similar work

MILD: Tractable Terrain Modeling for Learning Improved Bipedal Locomotion on Deformable Surfaces

Multi-Terrain Mastery: A Comprehensive Controller for Bipedal Locomotion

Robust bipedal locomotion on flowable slopes via foot-driven terrain manipulation