Advanced Locomotion for Quadruped Robots Through Reinforcement Learning

Search and Rescue (SAR) operations face significant challenges in hazardous environments where rescuers are at high risk. In post-disaster scenarios, such as building collapses, chemical spills, or earthquake-affected zones, access to critical areas is often delayed for safety reasons. Deploying robots equipped with advanced sensors can provide essential real-time data to rescue teams, reducing risks and improving decision-making. Quadruped robots are particularly suited for SAR operations due to their ability to traverse difficult terrain, offering a clear advantage over wheeled or tracked robots. These robots can gather crucial information via cameras, LIDAR, and thermal imaging, aiding in locating survivors and assessing structural damage. For quadruped robots to operate effectively in such challenging conditions, robust control strategies are essential. While Model Predictive Control (MPC) has been widely used for precise locomotion, it struggles in unpredictable environments due to high computational demands. Reinforcement Learning (RL) is emerging as a more adaptive and computationally feasible alternative, enabling robots to learn and adjust to complex, dynamic situations typical of SAR missions. This project proposes a structured methodology to enhance quadruped robot locomotion in SAR environments, focusing on RL-based control strategies. The proposed framework is divided into three phases: training, testing, and deployment. In the training phase, robots learn to walk using Deep RL, specifically the Proximal Policy Optimization (PPO) algorithm, which was optimized through a hyperparameter study. Results showed an improvement of 8.85% in general locomotion and 16.73% in obstacle avoidance by fine-tuning the discount rate and entropy coefficient. The testing phase integrates these trained policies into the Robot Operating System (ROS) for evaluation, leveraging simulated environments in Gazebo with sensor data fusion from two Intel RealSense cameras. Finally, the deployment phase involved building a physical test circuit and conducting real-world tests with the Aliengo robot. The established baseline for basic locomotion using the manufacturer’s control system provides a qualitative reference for future RL-based strategies. The findings of this research highlight the potential of RL to significantly improve robot performance in SAR tasks, with further development needed to enhance robustness and real-world deployment.