Experiments Towards Robotic Learning by Imitation

John Demiris

Abstract

This paper presents a modular robotic architecture designed to investigate the mechanisms of learning by imitation, bridging the gap between biological ethology and artificial intelligence. By employing a "teacher-learner" paradigm within a maze environment, the system identifies significant events to construct a functional mapping of actions to environmental states. This approach allows a learner robot to move beyond simple mimicry, enabling it to navigate novel environments independently by applying generalized principles extracted from the teacher's behavior.

1. Introduction

In the mid-1990s, "imitation" emerged as a focal point in ethology, yet the field lacked consistent terminology and a mechanistic understanding of how animals actually learn from one another. This research posits that the most effective way to validate biological theories of imitation is through synthetic modeling—specifically, by building robotic systems that implement these theories.

The primary challenge addressed in this work is the distinction between simple mimicry and functional learning. While a robot might easily reproduce a sequence of movements (akin to a dance), such behavior does not prove the robot has "learned" the task. The goal of this architecture is to demonstrate that a robot can observe a teacher’s strategy and apply the underlying principles to solve different, unseen problems.

2. A Modular Architecture for Imitative Learning

The proposed system is built upon five largely independent modules, each responsible for a specific facet of the imitation process:

  1. Maze Negotiation: This module defines the teacher's strategy, which can range from basic dead reckoning to sophisticated wall-following techniques.
  2. Teacher Following: The learner must maintain proximity to the teacher. This involves tracking algorithms that can filter out "loitering" or irrelevant movements to ensure the learner stays focused on the task-relevant path.
  3. Significant Event Perception: This serves as the cognitive trigger for learning. Rather than recording every movement, the robot identifies specific moments—such as a 90-degree turn or an external signal—that indicate a decision has been made.
  4. Environment Perception: At the moment a significant event is detected, the robot captures the state of the environment using its sensory suite, which may include ultrasonic, infra-red, or vision sensors.
  5. Derivation of Appropriate Action: This is the core learning engine. It creates an association between the perceived environmental state and the action performed by the teacher, potentially implemented through rule-based logic or connectionist (neural network) models.

3. The "Significant Event" Filter

A critical insight of this architecture is the rejection of continuous recording. In biological systems, imitation is rarely a frame-by-frame copy of an actor's movements. By focusing on "significant events," the robot filters out environmental noise and focuses on the decision-making nodes of a task. This allows the robot to build a sparse but highly functional map of the environment, associating specific sensory signatures with necessary directional changes. This filtering process is essential for moving beyond rote memorization toward true behavioral acquisition.

4. Experimental Methodology and Results

The architecture was initially tested within a simulated maze environment. In these trials, the learner robot followed a teacher through a training maze, extracting rules based on the teacher's navigation strategy.

The primary benchmark for evaluating the system was the robot's performance in a novel maze. If the robot could navigate a previously unseen environment using the rules it derived during the training phase, it proved that the system had achieved functional generalization rather than rote memorization.

Results from the simulation were successful, showing that the learner could traverse increasingly complex mazes. Following these results, the architecture began its transition to physical mobile robots equipped with ultrasonic and infra-red sensors to test its robustness against real-world sensor noise and physical constraints.

5. Conclusion and Future Work

This research demonstrates that robotics can serve as a powerful tool for achieving "explanatory adequacy" in biological hypotheses. By formalizing the process of imitation into a modular architecture, we can move closer to understanding how complex behaviors are acquired.

While the current results are promising for simple learning situations, future work will focus on increasing the complexity of individual modules—such as incorporating advanced computer vision—and applying the architecture to more dynamic tasks, such as obstacle avoidance in changing environments.

Summary updated

Abstract

Learning by imitation is a form of learning, which despite the fact it has been widely studied by ethologists, has not been fully understood yet. In fact, there is a considerable disagreement even on the terminology used despite attempts to clarify it (Davis,1973; Galef, 1988;). We believe that by building robots which instantiate the mechanisms hypothesised to underlie these types of behaviour, those mechanisms will be illuminated with explanatory adequacy. Our investigation into imitative learning begins with the construction of an appropriate experimental testbed and the design of a suitable architecture which would enable one robot (the learner) to imitate another one (the teacher) which is performing a task, and learn to perform the task while imitating the teacher. Initially, the idea was to have one robot to perform a sequence of moves (“dance”), while the other robot would learn dancing by imitating the first one. Such an experiment however, does not satisfy our need to be able to evaluate whether the robot has actually learned anything. Instead, we choose to have as a testbed a maze, where the robot learner would imitate the actions that the robot teacher is performing during its maze negotiation strategy. After the learning phase, we could ask the second robot to navigate itself through a new (different) maze, which would provide us with a more evident demonstration that learning has taken place.
BibTeX
@inproceedings{demiris1994experiments,
  author = {John Demiris},
  title = {{Experiments Towards Robotic Learning by Imitation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure