Reproduce “DeepSym: Deep Symbol Generation and Rule Learning from Unsupervised Continuous Robot Interaction for Planning” on OpenResearch

About this paper

Alper Ahmetoglu, M. Yunus Seker, Justus Piater, Erhan Oztop, Emre Ugur

DeepSym is a deep neural architecture that autonomously discovers action-grounded, discrete object and effect categories and learns probabilistic rules from unsupervised robot interaction, enabling high-level symbolic planning. It achieves 95% planning success and 70% execution success in complex manipulation tasks, demonstrating superior performance over alternative symbol learning approaches.

Key insights

• Action-effect regulated learning generates functional object and relational symbols that are inherently relevant for predicting interaction outcomes.
• Employing discrete units directly within the deep network's bottleneck during training yields superior symbol categorization and planning performance compared to post-clustering continuous latent spaces.
• The framework supports progressive learning, enabling the system to build richer abstractions and discover new task-relevant features as its interaction repertoire expands.

Summary by alphaXiv.