Navigation on Complex Terrains for the ICRA 2025 Quadruped Robot Challenges

As part of the Robodog Flagship Project, the objective is to compete in the autonomous navigation track of the ICRA 2025 Quadruped Robot Challenges (QRC) [1], set to take place next May in Atlanta, USA. The aim is to develop a robust navigation system that enables a small-scale resource-constrained quadrupedal robot to autonomously navigate complex 3D terrains, building upon an existing reinforcement learning (RL)-based locomotion controller. This project will design and implement a complete real-time hierarchical navigation system, incorporating terrain mapping, global and local planning, achieving effective navigation in highly challenging environments with obstacles akin to those in the competition. Both classical and RL planners will be explored to determine the most effective approach. The NVIDIA Isaac Sim/Lab frameworks will be used to simulate both the robot's locomotion and its perception through sensors including novel depth cameras and 3D Lidar. Realistic simulation environments will be used for both RL training and optimizing the navigation system, before deploying in the real world on a custom system built on top of a Unitree Go1 robot. Recent advancements in both classical [2] and learning-based planners [3,4], as well as hybrid approaches [5], show significant potential. Most of the research either focuses on a single component or uses a bigger robotic platform with more resources. The integration of real-time terrain mapping and traversability estimation [6], already employed by the current locomotion controller, shows potential for deploying a full navigation and locomotion pipeline in resource-constrained systems. This project offers the opportunity to work on various components of the navigation system, including perception, planning and control. It is structured to allow the exploration of both established and novel methods for improving robot navigation.

As part of the Robodog Flagship Project, the objective is to compete in the autonomous navigation track of the ICRA 2025 Quadruped Robot Challenges (QRC) [1], set to take place next May in Atlanta, USA. The aim is to develop a robust navigation system that enables a small-scale resource-constrained quadrupedal robot to autonomously navigate complex 3D terrains, building upon an existing reinforcement learning (RL)-based locomotion controller.

This project will design and implement a complete real-time hierarchical navigation system, incorporating terrain mapping, global and local planning, achieving effective navigation in highly challenging environments with obstacles akin to those in the competition. Both classical and RL planners will be explored to determine the most effective approach.

The NVIDIA Isaac Sim/Lab frameworks will be used to simulate both the robot's locomotion and its perception through sensors including novel depth cameras and 3D Lidar. Realistic simulation environments will be used for both RL training and optimizing the navigation system, before deploying in the real world on a custom system built on top of a Unitree Go1 robot.

Recent advancements in both classical [2] and learning-based planners [3,4], as well as hybrid approaches [5], show significant potential. Most of the research either focuses on a single component or uses a bigger robotic platform with more resources. The integration of real-time terrain mapping and traversability estimation [6], already employed by the current locomotion controller, shows potential for deploying a full navigation and locomotion pipeline in resource-constrained systems.

This project offers the opportunity to work on various components of the navigation system, including perception, planning and control. It is structured to allow the exploration of both established and novel methods for improving robot navigation.

Develop and deploy a robust, real-time navigation system for a quadrupedal robot, enabling it to autonomously navigate complex 3D terrains and compete successfully in the ICRA 2025 Quadruped Robot Challenges. **Tasks** - Review literature on quadrupedal robot navigation in uneven terrains. - Familiarize with the RL controller and the simulation/training environment. - Extend the simulation to include additional sensors and a navigation framework. - Select the most promising approaches from the literature. - Implement and integrate selected baselines with RL controller. - Propose and implement improvements, and fine-tune for real-time operation. - Test system in simulation and real world. - Fine-tune system for the ICRA Quadrup Robot Challenges. - Evaluate the proposed system and baselines. **Prerequisites** - Solid programming experience (Python and C++). - Experience with ROS. - Experience with robotics navigation and RL. - Optional: experience in robotics simulators. **Character** - 10% Literature study - 20% Implementation of baselines - 20% Implementation of improvements - 20% Experiments - 20% Evaluation and Analysis - 10% Report and Presentation **References** [1]: A. Jacoff et al., "Taking the First Step Toward Autonomous Quadruped Robots: The Quadruped Robot Challenge at ICRA 2023 in London [Competitions]," in IEEE Robotics & Automation Magazine, vol. 30, no. 3, pp. 154-158, Sept. 2023, doi: 10.1109/MRA.2023.3293296. [2] Wellhausen, Lorenz, and Marco Hutter. "Artplanner: Robust legged robot navigation in the field." arXiv preprint arXiv:2303.01420 (2023). [3] Rudin, Nikita, et al. "Advanced skills by learning locomotion and local navigation end-to-end. In 2022 IEEE." RSJ International Conference on Intelligent Robots and Systems (IROS). [4] Yang, Ruihan, et al. "Learning vision-guided quadrupedal locomotion end-to-end with cross-modal transformers." arXiv preprint arXiv:2107.03996 (2021). [5] Lee, Joonho, et al. "Learning robust autonomous navigation and locomotion for wheeled-legged robots." Science Robotics 9.89 (2024): eadi9641. [6] Miki, Takahiro, et al. "Elevation mapping for locomotion and navigation using gpu." 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022.

Develop and deploy a robust, real-time navigation system for a quadrupedal robot, enabling it to autonomously navigate complex 3D terrains and compete successfully in the ICRA 2025 Quadruped Robot Challenges.


**Tasks**

- Review literature on quadrupedal robot navigation in uneven terrains.
- Familiarize with the RL controller and the simulation/training environment.
- Extend the simulation to include additional sensors and a navigation framework.
- Select the most promising approaches from the literature.
- Implement and integrate selected baselines with RL controller.
- Propose and implement improvements, and fine-tune for real-time operation.
- Test system in simulation and real world.
- Fine-tune system for the ICRA Quadrup Robot Challenges.
- Evaluate the proposed system and baselines.



**Prerequisites**

- Solid programming experience (Python and C++).
- Experience with ROS.
- Experience with robotics navigation and RL.
- Optional: experience in robotics simulators.


**Character**

- 10% Literature study
- 20% Implementation of baselines
- 20% Implementation of improvements
- 20% Experiments
- 20% Evaluation and Analysis
- 10% Report and Presentation




**References**

[1]: A. Jacoff et al., "Taking the First Step Toward Autonomous Quadruped Robots: The Quadruped Robot Challenge at ICRA 2023 in London [Competitions]," in IEEE Robotics & Automation Magazine, vol. 30, no. 3, pp. 154-158, Sept. 2023, doi: 10.1109/MRA.2023.3293296.

[2] Wellhausen, Lorenz, and Marco Hutter. "Artplanner: Robust legged robot navigation in the field." arXiv preprint arXiv:2303.01420 (2023).

[3] Rudin, Nikita, et al. "Advanced skills by learning locomotion and local navigation end-to-end. In 2022 IEEE." RSJ International Conference on Intelligent Robots and Systems (IROS).

[4] Yang, Ruihan, et al. "Learning vision-guided quadrupedal locomotion end-to-end with cross-modal transformers." arXiv preprint arXiv:2107.03996 (2021).

[5] Lee, Joonho, et al. "Learning robust autonomous navigation and locomotion for wheeled-legged robots." Science Robotics 9.89 (2024): eadi9641.

[6] Miki, Takahiro, et al. "Elevation mapping for locomotion and navigation using gpu." 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022.

Davide Plozza (davide.plozza@pbl.ee.ethz.ch), Paul Joseph (josephp@student.ethz.ch)

Davide Plozza (davide.plozza@pbl.ee.ethz.ch), Paul Joseph (josephp@student.ethz.ch)