Publicação
Leveraging intrinsic motivations in robotics: empowerment
| Resumo: | This work investigates empowerment as an intrinsic motivation for enhancing autonomy and adaptability in robotic systems. Empowerment, defined as the informational capacity between a robot’s actions and its subsequent sensory states, emerges as an effective alternative to conventional external rewards, naturally fostering adaptive, safe, and self-sustaining behaviors. The research implements computational approaches to measure empowerment within realistic simulated environments using platforms such as MuJoCo and PyBullet. Methods include finite-difference approximations, optimization via the Blahut-Arimoto algorithm, integration with reinforcement learning, and notably, the application of the sensitivity based empowerment for efficient computation of empowerment in dynamical systems. These methods address challenges related to computational complexity and nonlinear dynamics. Experimental results demonstrate that empowerment-guided robots successfully perform complex tasks like balancing an inverted pendulum, strategic positioning in robotic soccer scenarios, and successful footstep planning for bipedal robots. The findings highlight empowerment’s effectiveness in ensuring robot safety and operability while acknowledging limitations in scalability and the necessity of integrating explicit external goals. Ultimately, empowerment is validated as a robust intrinsic criterion for robotic decision-making, offering substantial promise for further application in autonomous robotics. |
|---|---|
| Autores principais: | Figueiredo, Roberto Ribeiro |
| Assunto: | Empowerment Intrinsic motivation Robotic systems Reinforcement learning Dynamical systems Mutual information Robotic autonomy Blahut-Arimoto |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Aveiro |
| Idioma: | inglês |
| Origem: | RIA - Repositório Institucional da Universidade de Aveiro |
| Resumo: | This work investigates empowerment as an intrinsic motivation for enhancing autonomy and adaptability in robotic systems. Empowerment, defined as the informational capacity between a robot’s actions and its subsequent sensory states, emerges as an effective alternative to conventional external rewards, naturally fostering adaptive, safe, and self-sustaining behaviors. The research implements computational approaches to measure empowerment within realistic simulated environments using platforms such as MuJoCo and PyBullet. Methods include finite-difference approximations, optimization via the Blahut-Arimoto algorithm, integration with reinforcement learning, and notably, the application of the sensitivity based empowerment for efficient computation of empowerment in dynamical systems. These methods address challenges related to computational complexity and nonlinear dynamics. Experimental results demonstrate that empowerment-guided robots successfully perform complex tasks like balancing an inverted pendulum, strategic positioning in robotic soccer scenarios, and successful footstep planning for bipedal robots. The findings highlight empowerment’s effectiveness in ensuring robot safety and operability while acknowledging limitations in scalability and the necessity of integrating explicit external goals. Ultimately, empowerment is validated as a robust intrinsic criterion for robotic decision-making, offering substantial promise for further application in autonomous robotics. |
|---|