The Path to AGI Goes through Embodiment

Cheston Tan, Shantanu JaiswalAAAI 2023 · DOI

TL;DR

This paper argues that while Large Language Models (LLMs) have made impressive strides, they alone are insufficient for achieving Artificial General Intelligence (AGI). The authors contend that embodiment – an agent's interaction with the physical world – is not only necessary for AGI but also the most convincing way to demonstrate it, especially given the lack of a definitive AGI test.

What Problem Does This Paper Address?

The rapid advancement of LLMs has sparked a debate: can these models, trained on vast amounts of text, eventually lead to Artificial General Intelligence (AGI)? While LLMs excel at language-based tasks and exhibit impressive reasoning capabilities, many researchers question if this is enough. The core problem this paper addresses is the definition and demonstration of AGI. Specifically, it tackles the question of whether current AI, primarily LLM-based, can truly be considered AGI, and if not, what crucial element is missing. The authors highlight the lack of a universally accepted "acid test" for AGI, making its recognition subjective and dependent on convincing demonstrations of capability.

Methods

This paper is a position essay, not a traditional research paper with novel methods or algorithms. Instead, it presents a philosophical argument supported by existing research and logical reasoning. The authors' "method" involves:

  • Defining Key Terms: They provide working definitions for AGI ("an autonomous machine exhibiting general-purpose learning and reasoning capabilities that equal or surpass human-level capabilities across diverse tasks and in a broad range of environments or situations") and embodiment ("an agent's engagement with a physical or virtual environment through sensorimotor interactions and experiences, mirroring a human's experiential understanding of the world").
  • Analyzing Existing AI Paradigms: They discuss the strengths and limitations of LLMs and deep reinforcement learning (RL) in the context of AGI.
  • Constructing an Argument: They build a case for embodiment by refuting the sufficiency of language-only systems and formal knowledge, and by emphasizing the subjective nature of AGI recognition.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • AGI is Subjective Without Embodiment: The lack of a definitive "acid test" for AGI means its recognition is currently subjective. Embodied demonstrations are proposed as the most convincing way to overcome this subjectivity.
  • Language is an Interpretation, Not Reality: LLMs capture knowledge through symbolic representation, which is an interpretation of reality. This interpretation is compressed and relies on unstated commonsense assumptions, making it inherently incomplete and potentially brittle compared to direct experience.
  • Embodiment is Crucial for General-Purpose Robots: For AGI to be useful in practical applications like general-purpose robots, it must be able to interact with and understand the physical world, its constraints, and its dynamics.
  • Abstract Planning Isn't Enough: Even high-level planning requires an understanding of practical and physical realities (e.g., overhead clearance, floor loading capacity) that can only be gained through embodied experience.
  • Real-World Interaction is the Ultimate Test: While superhuman performance on intellectual tasks or broad knowledge is impressive, it's not sufficient for convincing AGI. Robust, continuous, real-world interaction and task performance are what will ultimately demonstrate AGI.
  • Children as a Benchmark: The paper draws a parallel to human children, who, despite limited formal knowledge, possess a remarkable ability to robustly interact with and "survive" in the real world, suggesting this embodied capability is a key indicator of general intelligence.

Key Results

As a position essay, this paper does not present quantitative results. Its "results" are conceptual:

  • Argument for Embodiment: The paper successfully argues that embodiment is a necessary component for AGI and the most convincing way to demonstrate it.
  • Critique of Language-Only AGI: It effectively critiques the notion that LLMs alone can achieve AGI, highlighting their limitations in capturing true understanding and real-world interaction.

Limitations and Future Work

  • Subjectivity of AGI: The paper acknowledges that the lack of a definitive test for AGI is a significant limitation, making the recognition of AGI inherently subjective.
  • "Something More" is Missing: The authors suggest that while components like advanced reasoning and broad knowledge are important, there's always been "something more" missing, which they identify as embodiment.
  • Future Work Implied: The paper strongly implies that future research should focus on developing embodied AI systems that can interact with and learn from the real world. This would involve creating more sophisticated robotic platforms and environments for testing and demonstrating AGI capabilities. The authors also suggest that researchers should consider how to bridge the gap between abstract knowledge and practical, embodied understanding.

Summary updated

Abstract

Recent advances in large language models have raised the question of whether these language models alone could lead to artificial general intelligence (AGI). In this short position essay, we argue that embodiment is not only required for achieving AGI, but also that embodiment is the key to convincingly demonstrate AGI capabilities. There is no single widely-accepted, objective test for AGI, so therefore whether a system has achieved AGI is a subjective judgement. We argue that a language-only system or one that cannot demonstrate success in the real world would not be convincing.
BibTeX
@inproceedings{tan2023path,
  author = {Cheston Tan and Shantanu Jaiswal},
  title = {{The Path to AGI Goes through Embodiment}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaaiss.v1i1.27485},
}
Figure