Humanoid Robot Infers Archimedes' Principle: From Cumulative Exploration to Abstraction of Underlying Physical Relations and Object Affordances
Authors: Bhat A. A., Mohan V., Sandini G., & Morasso P.
Journal: Journal of the Royal Society Interface
Tags: cognitive, robotics, causal, exploration, episodic-memory, tool-use, learning
Link: URL
Abstract:
Reenacting Aesop's Crow and Pitcher on a humanoid, we propose a neural architecture that encodes sensorimotor interactions into episodic traces and applies four learning rules (elimination, growth, uncertainty, status quo) to extract causal relations. The robot's predictions for novel objects converge to Archimedes' law.
Methodology:
An episodic memory network stores one-shot object-water interactions. Learning rules compare predicted to recalled outcomes to adjust feature weights. Generalization is tested on unseen objects; the abstracted model matches expected fluid displacement.