29 Mar 2021
16:00 Master's Defense Fully distance
Theme
Object Affordance Learning Through Robotic Interaction
Student
Renan Lima Baima
Advisor / Teacher
Esther Luna Colombini
Brief summary
Object affordance learning is the ability to process information about objects and how to use them. Embedding this knowledge in robots is an essential step for the development of intelligent and truly autonomous agents. Many researchers propose the development of specialized agents, networks of classifiers or advanced solutions for very specific scenarios, all of which demand high computational power. However, these approaches do not allow agents to acquire the necessary level of abstraction required by the complex environments in which they will be inserted. This work proposes the development of an incremental framework for learning affordances for robotic manipulation. The proposed approach was implemented as a reinforcement function in a network with a Soft Actor-Critic structure and trained in simulation with a humanoid robot. Among three different affordance complexities (touching, picking up and lifting the object), the results show a rate of up to 95% correct in the best scenario, with the agent properly performing all the desired actions. It also achieved the most complex affordance expected, which was not previously possible with the use of the goal-based strategy. In addition, the trained policy had a success rate in learning affordances of up to 68,5%, even when applied to previously untrained objects on the agent's scene. Such results suggest that an abstraction of movements, based on the approach of learning incremental affordances, may have been achieved.
Examination Board
Headlines:
| Esther Luna Colombini | IC / UNICAMP |
| Plinio Thomaz Aquino Junior | UNIFEI |
| Paula Dornhofer Paro Costa | FEEC / UNICAMP |
Substitutes:
| Sandra Eliza Fontes de Avila | IC / UNICAMP |
| Ricardo Ribeiro Gudwin | FEEC / UNICAMP |