DTA

Corso Ordinario Secondo Livello

VICARI, ANDREA

etd-10212024-203745

Energy-Aware RL-Trained Policy for Energy-Efficient Quadruped locomotion

Classe Scienze Sperimentali

INGEGNERIA - INGEGNERIA

relatore Prof.ssa MENCIASSI, ARIANNA
relatore Prof. AVIZZANO, CARLO ALBERTO

• reinforcement learning
• robotics

09/12/2024;

parziale

This thesis presents a novel approach to improve energy efficiency in legged robots by using Long Short-Term Memory (LSTM) neural networks to model actuator energy consumption. This new method outperforms traditional models by offering more accurate predictions, which are key to reducing energy usage. The method was integrated into the Isaac Gym framework, where reinforcement learning algorithms were used to train an energy-aware policy. Applied to real robots, this policy reduced energy consumption by around 25%, with smoother and quieter movement. Initially implemented on the ANYmal C robot, the method has been also tested on ANYmal D, a new generation quadruped, achieving similar results.

Nome file	Dimensione
Ci sono 1 file riservati su richiesta dell'autore. Contatta l'autore