• reinforcement learning
• robotics

This thesis explores an innovative approach to increase energy efficiency in legged robots, an area that has received a lot of attention in robotics research. The work introduces a new method for modeling the energy consumption of robot actuators using Long Short-Term Memory (LSTM) neural networks. This technique marks a significant shift from traditional methods by providing more accurate predictions of energy needs, crucial for developing strategies to reduce energy usage in robotic systems. The effectiveness of this method is evaluated against existing models and validated through experiments on actual hardware using certified measuring tools. The advanced prediction method is integrated into the Isaac Gym framework to train a policy that optimizes energy consumption using RL algorithms.
The practical application of this energy-optimized policy on real robotic systems led to significant improvements. The overall energy consumption of the robot was reduced by approximately 25%, with the energy usage of the knee actuators cut by half. Moreover, the robot exhibited smoother and quieter movement. The research initially involved the ANYdrive C actuator of the ANYmal C quadruped and was later extended to the newer ANYdrive D actuator, yielding comparable results.