Autonomous Navigation and Walking Wheel-Leg Robot

Background Introduction

The acceleration of urbanization has posed significant challenges for supply chain logistics, especially for last-mile delivery. As traffic pressure increases and the demand for faster delivery services rises, particularly with complex routes indoors and on city streets, traditional wheel-based robots struggle to overcome obstacles, and leg-only systems cannot achieve the required speed and efficiency. For instance, the ANYmal robot, while possessing some mobility, only achieves half the average human walking speed and has limited battery life. Therefore, developing a robot system capable of both efficient movement on flat surfaces and overcoming obstacles has become a primary research direction.

This paper focuses on wheel-leg robots, combining the advantages of wheels and legs to achieve high-speed travel on medium-strength surfaces over long distances while maintaining flexibility on complex terrains.

Paper Source

This paper was authored by Joonho Lee, Marko Bjelonic, Alexander Reske, Lorenz Wellhausen, Takahiro Miki, and Marco Hutter. The authors are from ETH Zurich, Swiss-Mile Robotics AG, and Neuromeka. The paper was published on April 24, 2024, in the journal “Science Robotics”.

Research Workflow and Details

Research Process

1. System Design and Development:

   • This paper developed a comprehensive system for wheel-leg robots, including adaptive motion control, localized navigation planning, and large-scale path planning.
   • A general motion controller was developed using model-free Reinforcement Learning (RL) technology and privileged learning.
   • To achieve efficient urban scene navigation, an integrated hierarchical RL framework was designed to enable high-speed effective navigation through complex terrains and various obstacles.

2. Experiments and Validation:

   • Autonomous navigation tasks were conducted in Zurich, Switzerland, and Seville, Spain, to verify the robustness and adaptability of the system.
   • Controlled training was conducted using digital twin models and real-time positioning with simulated data.

Main Experimental Content

1. Robot Hardware Configuration:

   • The robot is equipped with various sensors, including Lidar, stereo cameras, 5G routers, and GPS antennas to support localization and dynamic path detection.
   • Real-time personnel tracking is achieved through high-frequency object detection, creating a buffer within a 20-meter range to enhance safety.

2. Navigation System:

   • The system sets up a global path and uses a highly adaptive controller to generate velocity target commands to guide the robot along the path.
   • The controller makes sensible travel decisions using hidden states generated by the lower-level controller, terrain height values, and previously visited position sequences.
   • Input data is processed using a combination of 1D and 2D convolutional neural networks with multilayer perceptrons for timely responses.

3. Motion Controller:

   • The lower-level controller, based on model-free RL technology, uses a recurrent neural network (RNN) to learn the robot’s smooth transition between walking and driving modes.
   • Performance is enhanced using privileged information during the training process, with the final policy relying only on raw data from the inertial measurement unit (IMU) and joint encoders.

Experimental Results

1. Large-Scale Autonomous Deployment:

   • The robot performed long-distance autonomous navigation tasks in Zurich and Seville, covering a total distance of 8.3 kilometers.
   • During the experiments, the robot demonstrated its ability to navigate through various obstacles and different terrains with an average speed of 1.68 m/s and a mechanical transport cost (Cotmech) of 0.16.

2. Local Navigation and Embedding:

   • In various navigation scenarios, the robot displayed its ability to detect obstructed paths, navigate through complex obstacles, and select appropriate gaits.
   • Dynamic obstacle recognition was achieved using cameras combined with human detection, ensuring safe avoidance of pedestrians.

3. Hybrid Movement:

   • The lower-level controller was tested on different terrains, displaying highly adaptive gaits and stable body posture control, reaching a maximum speed of 5.0 m/s.
   • On steep slopes, stairs, and other complex terrains, the robot smoothly transitioned between crawling and driving modes.

Research Conclusions

This paper demonstrates through the design and validation of an autonomous navigation system for wheel-leg robots the potential for efficient and robust autonomous navigation in complex urban environments. The research results not only validate the effectiveness of the system architecture and control strategy but also highlight the promising prospects of wheel-leg robots replacing human labor for last-mile delivery.

Research Highlights

1. Novel System Integration Architecture: The seamless integration of adaptive motion control, localized navigation planning, and large-scale path planning significantly enhances the robot’s navigation capability in complex environments.

2. Hybrid Movement Control: By combining model-free RL technology with privileged learning, an efficient and robust motion controller was developed, enabling the robot to switch gaits and maintain efficient movement across various complex terrains.

3. Practical Verification: Large-scale autonomous navigation experiments conducted in Zurich and Seville validated the system’s adaptability and robustness, providing vital practical references for future applications.

The study shows that wheel-leg robots hold great potential in addressing urban logistics last-mile delivery challenges and offers new directions and technological support for the development of autonomous navigation and intelligent mobility in robotics.