Reproduce “A Survey of Robotic Language Grounding: Tradeoffs between Symbols and Embeddings” on OpenResearch

About this paper

Vanya Cohen, Jason Xinyu Liu, Raymond Mooney, Stefanie Tellex, David Watkins

The authors survey robotic language grounding as a spectrum from commands mapped to human-defined structures to learned language-and-perception representations that produce actions. They classify prior systems and show that many combine both approaches, including systems with fixed skill vocabularies and learned components. The survey describes conditional tradeoffs but reports no controlled comparison; robots also lack capabilities for many requested tasks.

Key insights

• Robotic language grounding approaches can be effectively categorized along a spectrum from human-defined formal representations (symbols) to learned high-dimensional vector embeddings.
• Each end of this spectrum presents distinct tradeoffs regarding precision, data efficiency, interpretability, generality, and scalability.
• Many successful contemporary robotic systems implicitly or explicitly combine elements of both symbolic reasoning and data-driven embedding approaches.

Summary by alphaXiv.