A Dynamical Systems Approach for a Learnable Autonomous Robot

Jun Tani, Naohiro FukumuraNeurIPS 1995

Traditional robot navigation typically relies on either a global coordinate map or explicit landmark detection. Both approaches have significant flaws: global maps are difficult to build autonomously, and landmark matching is brittle—if a robot misidentifies a landmark or gets "lost," it often cannot recover.

This paper proposes a paradigm shift in robot navigation, moving away from global mapping toward an "internal" dynamical systems approach. By utilizing a Recurrent Neural Network (RNN) to process local sensory data, a robot can learn complex, multi-step paths (such as a figure-eight) as stable "attractors." This allows the robot to recover autonomously even if it gets lost or encounters environmental noise, effectively addressing the non-Markovian nature of navigation where a single sensor snapshot isn't enough to decide the next move.

Methodology: A Two-Level Neural Architecture

The researchers implemented their system on a YAMABICO mobile robot using a two-level architecture:

  1. The Control Level (Fixed): A low-level reactive system that uses laser range data to avoid obstacles and instinctively head toward open spaces.
  2. The Navigation Level (Learned): A higher-level system that makes decisions at "branching points" (e.g., intersections) based on learned experiences.

To bridge these levels, the system combines two distinct neural network techniques to process sensory input and maintain a sense of history.

Neural Architecture

First, a Kohonen Network (Self-Organizing Map) compresses raw 24-dimensional laser range profiles (sampled over a 160-degree arc) into a manageable 3D representation. Second, an Elman-type Recurrent Neural Network (RNN) processes this compressed data. Unlike standard feed-forward networks, this RNN features "context units" that store a memory of past states. This memory allows the robot to distinguish between two identical-looking hallways based entirely on its previous trajectory.

Experimental Setup

Instead of relying on standard public datasets, the study used real-world experimental data collected directly from the YAMABICO robot. The training data was generated by a human trainer guiding the robot through a specific workspace from 10 different starting locations. The primary task was a "cyclic routing" challenge where the robot had to repeatedly navigate a figure-eight ('8') followed by a single loop ('0').

Key Insights: Navigation as an Attractor

The core innovation of this paper is framing navigation as an embedding problem. Instead of programming a rigid "if-then" script, the goal is to train the RNN so that the desired path becomes a "global attractor" in the robot's internal state space.

Attractor Dynamics
  • Stable Orbits: The "8-0" loop becomes a stable orbit (a limit cycle). If the robot starts in a random corner of the room, it wanders until it "catches" the rhythm of the learned path and then stays locked into that behavior.
  • Context Trumps Noise: The researchers found that while raw sensory inputs were often noisy or "chattering" at branching points, the RNN's internal context units remained remarkably stable. The internal memory proved more reliable than the physical sensors of the robot.
  • Self-Correction: Because the task is embedded as a global attractor, the robot exhibits "structural stability." It doesn't need an explicit "re-localize" command; the underlying mathematics of its state-space transitions naturally pull it back onto the correct path if it is nudged off-course.

Results and Limitations

The experiments demonstrated successful convergence: regardless of the starting position, the robot consistently found its way to the desired "8-0" looping sequence. It successfully navigated even when sensory inputs were unstable or when the number of branching points varied due to environmental noise. Analysis of the RNN confirmed that the "context units" clearly differentiated between the two different loops of the task, even when the local laser images looked nearly identical.

However, the authors note several limitations. In highly unpredictable or "rugged" environments, the structural stability can break down if the "attractor" isn't strong enough to hold against the chaos. Furthermore, the current model learns a specific "state-action map" for a fixed task. Future work will need to explore how robots can build "analogical models" of the environment to allow for flexible goal-setting without requiring a complete retraining of the network. Finally, scaling this approach from a controlled workspace to a large, dynamic office or outdoor environment remains a significant challenge.

Summary updated

BibTeX
@inproceedings{tani1995dynamical,
  author = {Jun Tani and Naohiro Fukumura},
  title = {{A Dynamical Systems Approach for a Learnable Autonomous Robot}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1995},
}
Figure