17 Mar 2025
09:00 Doctoral defense IC3 Auditorium
Topic
A Motivation-Driven Incremental Learning Framework for Robotics
Student
Letícia Mara Berto
Advisor / Teacher
Esther Luna Colombini - Co-advisor: Ricardo Ribeiro Gudwin
Brief summary
The integration of robots into daily activities is growing, requiring their adaptation to increasingly complex environments. This adaptation requires the development of learning capabilities through interaction with the environment and other agents, allowing the execution of tasks beyond the programmed instructions. At the heart of this process is cognition, the ability to transform information from diverse sources into practical knowledge through interconnected cognitive functions. %These functions must act cohesively, but their emergence probably depends on fundamental conditions.
Our research is grounded in cognitive and developmental robotics, inspired by human development processes, and contributes to the understanding of how robots integrate motivation, decision-making, and adaptability to operate autonomously and interact meaningfully with humans and other robots in dynamic environments. Initial experiments focused on the robot's interaction with the environment, while later studies incorporated social interaction. One of the central challenges in robotics is replicating decision making. So we adapted a human experiment, involving uncertainty, rewards and punishments, to the robotic context. Using reinforcement learning, we train simulated robots to maintain homeostasis. The results indicated that robots that adopted long-term strategies not only survived longer, but also better satisfied various needs.
In addition to meeting basic needs, humans develop preferences influenced by perceived pleasure, adding complexity to decision making. To model this, we extend Hull’s motivational theory by incorporating hedonic dimensions that allow robots to balance needs and preferences according to context. We trained three agents with different metabolic rates in different environments. Our results showed that agents adapted more effectively when environmental conditions were aligned with their metabolic needs. Furthermore, the introduction of pleasure into the motivational model influenced behavior, especially in agents with moderate metabolic rates. Under survival pressures, agents prioritized immediate needs over pleasure, highlighting the dynamic interaction between motivation, pleasure and environmental context in decision-making.
To explore human-robot interaction, we developed a cognitive architecture inspired by the needs of young children, incorporating three needs: learning, interaction, and recharging. Adjusting the importance of these needs, we created two robot profiles: Playful and Social. Experiments have shown that changes in the importance of needs resulted in different behaviors in robots and human perceptions. Participants adapted their interactions to the robots’ behaviors, assigning characteristics aligned with the profiles, even without prior knowledge.
Finally, we examine interactions between two robotic agents with varying needs, using theory of mind to estimate and respond to each other's demands.
Examination Board
Headlines:
| Esther Luna Colombini | IC / UNICAMP |
| João Eduardo Kogler Junior | PSI/USP |
| Diana Francisca Adamatti | C3 / FURG |
| Sandra Eliza Fontes de Avila | IC / UNICAMP |
| Carlos Henrique Costa Ribeiro | COMP/ITA |
Substitutes:
| Julio Cesar dos Reis | IC / UNICAMP |
| Rodrigo da Silva Guerra | C3 / FURG |
| Reinaldo Augusto da Costa Bianchi | FEI |