Instructo-Soar: Learning from Interactive Natural Language Instructions

AAAI 1993

TL;DR

This paper introduces Instructo-Soar, a system that learns new robotic procedures directly from interactive, imperative natural language instructions. It addresses the challenge of teaching AI agents complex tasks by enabling them to learn not just sequences of actions but also the underlying goal concepts and generalizable procedures, mimicking human learning.

What Problem Does This Paper Address?

Teaching artificial intelligence agents to perform complex tasks is a significant challenge. While many AI systems can be programmed or learn through demonstration, learning from natural language instructions, a fundamental human learning mechanism, has seen limited exploration. This paper tackles the problem of enabling an AI agent to acquire new skills and procedures by understanding and acting upon imperative natural language commands given in an interactive, situated context. This means the agent is actively engaged in a task and receives instructions as needed, rather than being given a static set of rules. The goal is to move beyond rote memorization towards genuine understanding and generalization of learned procedures.

Methods

Instructo-Soar employs a multi-stage learning process. Initially, it learns new procedures through rote memorization of the instruction sequences. Following the first execution of a newly learned procedure, the system uses a "difference-of-states" heuristic to infer the goal concept. This means it identifies what has changed from the initial state to the final state as the intended outcome.

For generalization, Instructo-Soar uses an "explanation-based" approach. The agent internally simulates the execution of each instruction to understand how it contributes to achieving the inferred goal. This process helps identify the crucial situational features and instruction components necessary for success. The system learns a hierarchy of complex operators from a small set of primitive ones, and its learning curve follows a power law of practice, indicating increasing efficiency with experience. Recent versions also allow for instructional feedback to refine the induced goal.

Datasets

Not discussed in the paper. The paper describes an application in a simple robotic domain and uses an example instruction scenario shown in Figure 1, but does not refer to specific, named datasets for training or evaluation.

Key Insights and Ideas

  • Learning from Imperative Language: The core insight is that AI agents can learn complex procedures by directly interpreting and acting on imperative natural language commands, a more intuitive and human-like learning method.
  • Beyond Rote Memorization: The system moves beyond simply remembering instruction sequences by inferring goal concepts and generalizing procedures through explanation-based learning.
  • Situated and Interactive Learning: Learning is most effective when the agent is actively engaged in a task (situated) and can request instruction as needed (interactive), mirroring real-world learning scenarios.
  • Goal Inference via State Differences: The "difference-of-states" heuristic provides a simple yet effective mechanism for inferring the intended goal of a procedure from observed state changes.
  • Power Law of Practice: The learning efficiency of Instructo-Soar follows a power law, suggesting that with more practice, the agent becomes significantly faster and more adept at learning new procedures.

Key Results

The paper demonstrates that Instructo-Soar can learn completely new procedures from interactive natural language instructions and extend its knowledge of previously known procedures to new situations. The system successfully learns a hierarchy of complex operators from a small set of primitives. The learning curve for acquiring new procedures closely matches the power law of practice, indicating efficient learning over time.

Limitations and Future Work

The authors acknowledge that this work represents "first steps towards our long-term goal of building general, instructable autonomous agents." While the paper demonstrates the system's capability in a simple robotic domain, the scalability and robustness of Instructo-Soar to more complex domains and a wider range of natural language instructions are implicit areas for future work. The paper also mentions the ability to incorporate instructional feedback, suggesting that further exploration of how to best utilize such feedback for more efficient and accurate learning is a promising direction.

Summary updated

BibTeX
@inproceedings{1993instructo,
  title = {{Instructo-Soar: Learning from Interactive Natural Language Instructions}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1993},
}
Figure