Talking with a Robot in English

This paper details the ENGROB system, an early attempt at enabling natural language communication between humans and a mobile robot. It explores the challenges of translating human language into a formal representation that a robot can understand and act upon, and discusses the implications for both AI and linguistics.

TL;DR

This paper presents ENGROB, a system designed to allow humans to communicate with a mobile robot using simple English. It details the architecture and components required for natural language understanding, translation into a formal logic, and execution of robot commands, highlighting the significant challenges in bridging the gap between human language and machine action.

What Problem Does This Paper Address?

The paper addresses the fundamental challenge of enabling natural language communication between humans and intelligent automata, specifically a computer-controlled mobile robot. The goal is to allow a human to engage in a dialog with the robot using simple English sentences, enabling the robot to understand commands, answer questions, and acquire information about its environment. This involves not just parsing sentences but also interpreting their meaning in the context of the robot's world model and capabilities, and then translating these into executable actions. The authors also posit that research in this area could significantly influence linguistic theory itself.

Methods

The ENGROB system is built around a six-component architecture:

  • Syntax Analyzer: Uses a transformational grammar for a subset of English (imperative, declarative, interrogative) to decompose complex sentences into simpler kernel sentences. This component is designed for efficiency, aiming to keep kernel sentence processing time constant.
  • Semantic Interpreter: Interleaves semantic productions with syntactic ones to translate well-formed English into first-order predicate calculus. This formal language serves as a "deep structure" representation.
  • Axiom Model: Stores information about the robot's environment, including geometric relationships, constraints on robot capabilities, and descriptive information from declarative sentences. This model is dynamically updated.
  • Inferential Component (QA3): A theorem-proving system based on resolution that deduces answers to questions or determines the feasibility of executing commands. It works by attempting to prove the negation of a hypothesis and looking for a contradiction.
  • Output-Sentence Generator: Uses a generative grammar, similar to the syntax analyzer, to produce English replies. These replies can be simple transformations of the input or more complex, incorporating semantic information.
  • Output-Action Generator: Translates imperative sentences and certain interrogative sentences into sequences of two-letter FORTRAN commands for the robot's primitive functions.

The core paradigm is to translate English into predicate calculus, perform inferences, and then generate an English output or a sequence of robot actions.

Datasets

Not discussed in the paper. The paper focuses on the system's architecture and capabilities rather than specific training datasets. The appendix provides sample English sentences and their translations into predicate calculus, serving as examples of the system's input and output.

Key Insights and Ideas

  • Predicate Calculus as a "Deep Structure": The paper advocates for using first-order predicate calculus as a robust internal representation for natural language meaning, bridging linguistic and non-linguistic information.
  • Syntax-Directed Interpretation: Integrating syntactic and semantic analysis within a common production framework is crucial for efficient language processing.
  • Transformational Grammar for Efficiency: Using transformations to simplify sentences before base component parsing can significantly improve efficiency.
  • The Influence of Robotics on Linguistics: The authors argue that the need to encode real-world knowledge and actions for robots will force linguistics to focus more on semantics and the relationship between language and reality, moving beyond purely grammatical aspects.
  • The Tacit Knowledge Gap: Humans often underestimate the vast amount of implicit knowledge (geometric, temporal, physical) required to understand and execute simple commands, a gap that AI systems must bridge.

Key Results

The paper presents a functional architecture for ENGROB and illustrates its operation with examples. While not providing quantitative performance metrics, it demonstrates the system's ability to:

  • Translate simple English declarative, interrogative, and imperative sentences into predicate calculus.
  • Process these translations to answer questions or generate commands for a robot.
  • Generate English responses based on the robot's state and inferred information.

The appendix lists 37 sample sentences, ranging from simple declarations like "All men are mortal" to complex descriptions and commands, showcasing the system's current translation capabilities.

Limitations and Future Work

The authors acknowledge that ENGROB is not yet fully implemented, and some aspects are speculative. Key limitations and areas for future work include:

  • Limited English Subset: The current grammar handles a subset of English; expansion is needed.
  • Efficiency of Theorem Proving: While QA3 is efficient, inferential requirements are expected to grow with more complex conversations, consuming a larger proportion of processing time.
  • Integration Complexity: Integrating the six components into a functioning whole within a complex time-sharing system, especially with hardware uncertainties, is a significant systemic challenge.
  • Theoretical Issues: Further research is needed on defining a canonical set of predicates for actions, refining QA3 heuristics, and exploring the adequacy of predicate calculus as a deep structure representation.
  • Real-World Interaction: The paper notes that the robot's physical capabilities (vision, manipulation) are still basic, and the interaction with the real world is a continuous area of development.
  • Ambiguity Resolution: The semantic component needs to participate more actively in parsing to resolve lexical and syntactic ambiguity with respect to the universe of discourse.

Summary updated

BibTeX
@inproceedings{1969talking,
  title = {{Talking with a Robot in English}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1969},
}
Figure