Competence Acquisition in an Autonomous Mobile Robot using Hardware Neural Techniques

Geoffrey Bruce Jackson, Alan F. MurrayNeurIPS 1995

Overview

Researchers at the University of Edinburgh have developed a custom VLSI chip, EPSILON II, that brings neural network processing directly to the "edge" of robotic sensors. By using a novel "instinct-rule" training methodology, they demonstrated that an autonomous robot can acquire basic mobility and obstacle avoidance skills in just a few training epochs. This work proves that analog hardware is uniquely suited for real-world sensor fusion, operating efficiently where traditional digital computers struggle.

The Analog-to-Digital Gap

Despite the theoretical success of neural networks, dedicated neural hardware has historically struggled to find a foothold. Fast, generic digital computers evolve so quickly that specialized digital accelerators are often obsolete by the time they hit the market.

However, the authors identify a specific gap: the interface between the messy, analog real world and high-level digital processing. Standard computers struggle with the overhead of converting and processing raw analog sensor data in real-time. This paper proposes that analog neural VLSI is the "niche technology" needed to bridge this gap, providing low-power, compact, and immediate processing of sensor data before it even reaches a main host processor.

Hardware Architecture: The EPSILON II System

To solve this bottleneck, the researchers developed a full-stack solution consisting of a custom chip and a supporting processor card.

Hardware Architecture of the EPSILON Processor Card
  • The EPSILON II Chip: A custom VLSI chip using "pulse stream" signaling. It represents neural states as pulses, combining the noise immunity of digital signals with the computational efficiency of analog circuits. It features a 32x32 synaptic matrix (1024 synapses) and can process both analog voltage levels and digital pulses simultaneously.
  • The EPSILON Processor Card (EPC): To make the chip usable, they embedded it on a bus-based card with an FPGA. As seen in the architecture diagram, the FPGA handles the routine tasks—weight refreshing, pulse encoding/decoding, and interfacing with the digital STE bus—allowing the neural chip to operate transparently to the host computer.

Instinct-Rule Training

Instead of traditional backpropagation, the robot uses a behavior-based approach driven by real-time environmental feedback.

Neural Control and Instinct-Rule Training Loop

The robot starts with random weights and a set of basic "instincts" (e.g., "Keep crash sensors inactive"). The neural network maps sensor inputs directly to motor outputs. If an instinct is violated (e.g., a crash sensor triggers), a monitor intervenes and the system tries the next strongest neural output. If that alternative action resolves the problem, the network's weights are updated to favor that mapping in the future.

Key Insights and Results

  • Somatic Tension: The authors introduced a biological analogy called "somatic tension." By summing all sensor inputs onto a single neuron with equal weights, they created a measure of total sensory "stress." If an action reduces this tension, the robot knows it’s moving in the right direction.
  • The "Edge" of Analog: The primary strength of analog VLSI isn't raw speed, but its ability to act as an "intelligent analog-to-digital converter." It pre-processes and interprets sensor data, presenting refined information to the host.
  • Noise as a Feature: In simulations, robots often get stuck in infinite oscillatory loops (e.g., vibrating back and forth in a corner). The authors found that in a real-world system, the inherent noise in analog sensors and mechanical asymmetries actually helps the robot "break out" of these loops, leading to more robust behavior than a perfect simulation would allow.
  • Rapid Learning: The robot achieved basic mobility competence (obstacle avoidance) in very few training epochs, starting from completely random weight initializations.
  • Successful Data Fusion: The system successfully mapped a complex mixture of analog sensor space to motor control outputs, demonstrating the flexibility of the EPSILON II architecture with significantly lower computational overhead for the host processor.

Limitations and Future Work

While the EPC supports multiple cards for larger networks, the current demonstration is limited to a relatively small 32x32 matrix. The authors acknowledge that this technology is not a replacement for general-purpose digital computing but is strictly optimized for the sensor-to-digital interface. Direct quantitative comparisons with previous robots were also difficult due to differences in physical chassis.

Moving forward, the researchers suggest that the next step is tackling more intensive navigational tasks by adding higher-level "navigational signal" rules to the existing instinct-rule framework.

Summary updated

BibTeX
@inproceedings{jackson1995competence,
  author = {Geoffrey Bruce Jackson and Alan F. Murray},
  title = {{Competence Acquisition in an Autonomous Mobile Robot using Hardware Neural Techniques}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1995},
}
Figure