Long-Distance Real-World Navigation of the Legged-Wheeled Robot Go2-W Using Deep Reinforcement Learning

arXiv:2606.21387 · cs.RO · Submitted 2026-06-19 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Long-Distance Real-World Navigation of the Legged-Wheeled Robot Go2-W Using Deep Reinforcement Learning".

Dev: Here's the gist of the paper. Legged-wheeled robots have been studied for their potential to combine the efficient flat-ground mobility of wheels with the rough-terrain capability of legs.

Rosa: This paper reports our effort to build a deep reinforcement learning (DRL) based locomotion controller and an autonomous navigation system for the commercially available legged-wheeled robot Go2-W, and to apply them to long-range autonomous navigation in a real environment.

Dev: System Overview The research presents an outdoor autonomous navigation system using the Go2-W robot and demonstrates successful long-distance travel in a real-world environment.

Rosa: The system combines original DRL-based methods for locomotion control and path-following control with conventional methods for localization and high-level planning.

Paper summary: Dev: So, looking at the title "Long-Distance Real-World Navigation of the Legged-Wheeled Robot Go2-W Using Deep Reinforcement Learning", it really summarizes what they did—they took an existing robot and applied deep reinforcement learning to make it navigate long distances in the real world.

Rosa: They extended locomotion control from quadrupeds to this legged-wheeled robot, found a way to deal with the heat concentration during wheeled movement, and put in a path-following controller that works with these new locomotion characteristics.

Taro: The main implication for us is that they proved that long-distance outdoor autonomous navigation is feasible with this specific robot configuration when you use this DRL approach.

Dev: It means the work on training locomotion policies to suppress heating, and how they analyzed the heat concentration mechanism, will be really useful for anyone working on legged-wheeled robots with similar setups.

Rosa: And their overall system design for navigation is also informative for how we approach these complex mobile platforms when you're planning long journeys outside the lab.

Conclusion: Rosa: So, this paper is about taking that Go2-W robot and using deep reinforcement learning to get it to travel long distances outside of a controlled lab setting in the real world.

Dev: Yeah, they actually built a whole system for that, combining their new control methods with path following and localization.

Rosa: It seems like they tackled a few big hurdles there, especially getting the robot to handle that heat problem during sustained movement.

Taro: I was looking at the results on the rough terrain policy, and it showed it could actually go up and down a single step that was twenty four centimeters high.

Dev: That's cool, but they also had this trade-off they mentioned—the conflict between keeping the robot cool and being able to traverse rough ground.

Rosa: Exactly, so the challenge isn't just making it move; it’s finding a balance where it doesn't overheat while still being capable of handling uneven surfaces.

Taro: I think that points toward future work where they try to learn just one policy that handles both thermal suppression and rough terrain capability at the same time.

Dev: And from an engineering standpoint, we need to know how they figured out that heat concentration issue; if you don't understand the mechanism, you can't build a reliable safety case for it.

Rosa: So what this means for us is that we have a working model for long-distance autonomous travel in real environments using this specific robot setup.

Dev: It opens up possibilities for deploying these legged-wheeled systems on longer missions where wheels and legs are both necessary tools.

Takaaki Matsuzawa, Kiyoshi Irie, Tomoaki Yoshida, Taro Suzuki, Yoshitaka Hara, Masahiro Tomono

Future Robotics Technology Center (fuRo), Chiba Institute of Technology

cs.RO

Submitted: 2026-06-19

Updated: 2026-11-13

Code: https://github.com/kiyoshiiriemon/simple

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 76/100

The gist: Legged-wheeled robots have been studied for their potential to combine the efficient flat-ground mobility of wheels with the rough-terrain capability of legs.

Key concepts

Deep Reinforcement Learning (DRL)
A machine learning approach where an agent learns how to perform tasks by interacting with an environment. The robot learns the best way to move and control itself through trial and error, using a policy trained via algorithms like PPO, allowing it to adapt its movements dynamically.
Locomotion Control Policy
The set of rules or learned behavior that dictates how the robot moves. The authors created different policies—stance, flat-terrain, and rough-terrain—specifically designed to manage the robot's movement while minimizing heat concentration in its hip joints during wheeled travel.
Domain Randomization
A training technique where random variations are intentionally introduced into the simulation data. This involves adding noise to sensor readings and introducing random action delays during training to help the learned policy generalize effectively from simulated conditions to the real world.
Heat Concentration Problem
A challenge where sustained wheeled locomotion causes excessive heat buildup in specific joints, particularly the hip joints. The DRL policy was specifically engineered to distribute this load more evenly across the robot's structure to suppress overheating and allow for longer travel times.

Terminology

Summary

Legged-wheeled robots have been studied for their potential to combine the efficient flat-ground mobility of wheels with the rough-terrain capability of legs. p. 0 This paper reports our effort to build a deep reinforcement learning (DRL) based locomotion controller and an autonomous navigation system for the commercially available legged-wheeled robot Go2-W, and to apply them to long-range autonomous navigation in a real environment. p. 0

System Overview

The research presents an outdoor autonomous navigation system using the Go2-W robot and demonstrates successful long-distance travel in a real-world environment. The system combines original DRL-based methods for locomotion control and path-following control with conventional methods for localization and high-level planning. p. 1 The hardware configuration includes the Unitree Go2-W, equipped with a Livox MID-360 3D Lidar and an NVIDIA Jetson Orin NX mounted in the rear expansion dock. p. 4 The software architecture is divided into an upper navigation layer and a lower locomotion control layer. p. 5

Locomotion Control Development

The authors extended a proprioception-only policy previously developed for quadruped robots to the 16-DoF legged-wheeled robot. They found that wheeled locomotion concentrates the load on the hip joints and causes heat concentration that hinders sustained travel, and obtained a policy that suppresses it by distributing the load. p. 0 Three types of policies are prepared: a stance policy, a flat-terrain locomotion policy, and a rough-terrain locomotion policy. p. 6 The stance policy is designed to maintain a stable standing posture. p. 8

Training Signal Construction

Locomotion policies were trained using an extended version of Legged Gym [16]. They employed the Asymmetric Actor Critic framework with Proximal Policy Optimization (PPO) as the learning algorithm. p. 7 Domain Randomization is used during training, adding Gaussian noise to sensor observations and introducing random action delays to bridge the sim-to-real gap. p. 7 The input to the policies consists of the velocity command and robot state observed through various proprioceptive sensors. p. 7

Performance and Results

The system was evaluated at the Tsukuba Challenge 2025, demonstrating that it can autonomously traverse an approximately 2.8 km route including sidewalks, a park, and stairs without stopping due to overheating. p. 0 The flat-terrain locomotion policy plateaued around 60 °C for the hottest motor during sustained locomotion, whereas the manufacturer-provided control reached 84 °C and shut down. p. 9 The rough-terrain policy was able to ascend and descend a single step up to a height of 24 cm. p. 9

Future Directions

Limitations include the trade-off between thermal suppression and rough-terrain traversal capability. Future work includes exploring methods to learn a single policy that achieves both thermal suppression and traversal capability, as well as automating terrain-dependent policy selection. Incorporating approaches such as occlusion-aware motion planning is also a direction for future work. p. 2 The study also suggests comparative evaluation of locomotion speed and energy efficiency across various environments with robots of other configurations. p. 2

Conclusion

In this paper, we developed a locomotion control and autonomous navigation system based on deep reinforcement learning for the commercially available legged-wheeled robot Go2-W. We extended locomotion control originally designed for quadruped robots to the legged-wheeled robot, obtained a locomotion policy that addresses the heat concentration problem during wheeled locomotion, and incorporated a path-following controller adapted to the locomotion characteristics of legged-wheeled mobility. p. 0 In the Tsukuba Challenge 2025, the robot completed an approximately 2. p. 08 km course including sidewalks, a park, stairs, and unpaved terrain, demonstrating the feasibility of long-distance outdoor autonomous navigation with a legged-wheeled robot. p. 0 The approach for training locomotion policies that suppress heating, the analysis of the heat concentration mechanism, and the navigation system design will be informative for work on legged-wheeled robots with similar configurations. p. 1

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Extended locomotion control to legged-wheeled robots by extending proprioception-only policies previously developed for quadruped robots to a 16-DoF system, allowing for stable operation on complex terrain.

  2. Development of a novel method for extending control developed for quadruped robots to legged-wheeled robots, addressing the differences in DRL problem formulations between the two robot types.

  3. Identification and analysis of heat concentration at specific joints during wheeled locomotion by identifying that wheeled locomotion concentrates the load on the hip joints and causes heat concentration that hinders sustained travel, leading to a policy that suppresses this by distributing load.

  4. Implementation of three distinct DRL policies—stance, flat-terrain, and rough-terrain—switched based on locomotion state commands, enabling the robot to transition between stable standing, efficient wheeled movement with foot lifting for heating reduction, and high-traversal rough terrain locomotion.

  5. Integration of a navigation system combining original DRL methods for locomotion control with conventional methods for localization and path-following control to enable long-distance autonomous navigation in real environments.

  6. Utilization of an Asymmetric Actor Critic framework with Proximal Policy Optimization (PPO) and Domain Randomization during training to enable direct sim-to-real transfer of proprioceptive policies trained in simulation to the physical robot.

  7. Deployment of a learning based odometry model that estimates relative movement from velocity command history, allowing for smooth path following even when ground contact state detection is difficult due to pneumatic tires.

  8. Enhancement of path-following control by training a data driven method where the locomotion policy is treated as a black box simulator to learn the robot’s short-term movement in response to velocity commands, resulting in smoother tracking compared to classical pure pursuit.

  9. Development of custom reward functions, including leg effort std and action curvature, which penalize imbalanced torque distribution and encourage smooth periodic motions, respectively, thereby promoting load equalization and reducing oscillatory behavior during locomotion.

  10. A navigation strategy that uses a combination of Lidar Inertial Odometry (LIO) with point cloud matching for self localization and a learning based odometry model to estimate local movement from velocity commands, providing robust pose estimation in outdoor environments.

Sources

Related papers