Acquisition of Automatic Activity through Practice: Changes in Sensory Input

Jack Gelfand, Marshall Flax, Raymond Endres, Stephen H. Lane, David HandelmanAAAI 1992

Introduction

This paper, originally presented at the 1992 AAAI proceedings, explores a biologically inspired approach to robotic control: the transition from "novice" visual guidance to "expert" proprioceptive automaticity. By using neural networks to learn from a visual "teacher," the authors demonstrate how robots can offload computationally expensive tasks to simpler, faster sensory-motor loops.

The Problem: Computational Bottlenecks in Robotics

In the early 1990s, roboticists faced a significant challenge: machine vision and global path planning were computationally expensive and often slow. Most systems relied on a single sensory modality, making them brittle. The authors draw a parallel to human motor skills: a novice basketball player must watch the ball to dribble, whereas an expert dribbles by "feel" (kinesthetic sense), freeing their eyes to scan the court. The goal of this research is to replicate this "automaticity"—the shift from high-level, visually-guided planning to low-level, reflexive response—in robotic systems to improve efficiency and reactivity.

Proposed Architecture: Hybrid Learning and Control

The authors propose a Hybrid Learning and Control System designed to facilitate this sensory transition.

The Hybrid Learning and Control System architecture, showing the interaction between visual and kinesthetic control loops.

The system consists of three primary components:

  1. Visual Planner/Controller: A high-level algorithmic system that uses vision to calculate trajectories and joint torques.
  2. Kinesthetic Controller (CMAC): A neural network based on the Cerebellar Model Articulation Controller (Albus, 1975). It acts as a function approximator, taking joint angles and velocities as input to predict required torques.
  3. Execution Monitor: A supervisor that compares the performance of the neural network against the visual controller. Once the neural network's error drops below a specific threshold, the monitor "switches" control from vision to internal sensors.

Experimental Scenarios

The system was evaluated using simulated 2D kinematic and dynamic models across three scenarios:

  • Obstacle Avoidance: A two-link arm learning to navigate around an object.
  • Ball Dribbling: A robotic arm interacting dynamically with a moving ball and an obstacle.
  • Calibration Compensation: A six-link "reacher" using vision to correct for random errors (noise) in its own joint sensors and limb lengths.

Key Results and Insights

The primary insight is the use of Vision as a Teacher. Instead of training a neural network from scratch via trial and error, the network "shadows" a traditional algorithmic controller until it can mimic the behavior autonomously. This leads to Sensory Offloading, where the visual system is eventually freed for other tasks without sacrificing motor performance.

Comparison of trajectories and relative error between visual control and neural network (N/N) control over multiple trials.

Key findings include:

  • Learning Curve: In obstacle avoidance, the CMAC's trajectory converged to the visual controller's path within a few dozen trials. As shown in the figure above, the relative error for the N/N control drops significantly as it learns the task, eventually reaching near-zero error and allowing for autonomous execution.
  • Dribbling Success: The dribbling robot successfully transferred control to the neural network after approximately 125 dribbles, learning a "control law" for dynamic interaction.
  • Error Correction: In the 6-link reacher experiment, the system demonstrated resilience. When sensors were perturbed, visual contribution spiked to 100% to correct the error, then dropped back toward zero as the CMAC learned the new "offset."

Conclusion and Future Work

The research demonstrates a viable path toward robotic automaticity. However, limitations remain, particularly regarding the dimensionality of the tasks (limited to 2D) and the simplicity of the visual models used. Future work suggests integrating more diverse sensory inputs (e.g., auditory or somatosensory) to allow the system to automatically select the most sensitive sensors for any given task.

Summary updated

BibTeX
@inproceedings{gelfand1992acquisition,
  author = {Jack Gelfand and Marshall Flax and Raymond Endres and Stephen H. Lane and David Handelman},
  title = {{Acquisition of Automatic Activity through Practice: Changes in Sensory Input}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1992},
}
Figure