Abstract
This paper describes the evolution of controllers for racing a simulated radio-controlled car around a track, modelled on a real physical track. Five different controller architectures were compared, based on neural networks, force fields and action sequences. The controllers use either egocentric (first person), Newtonian (third person) or no information about the state of the car (open-loop controller). The only controller that was able to evolve good racing behaviour was based on a neural network acting on egocentric inputs.
Introduction
While car racing is a multi-billion dollar industry, creating autonomous drivers that handle complex vehicle dynamics—such as momentum and skidding—remains a significant challenge. Most evolutionary robotics research utilizes simple models like the Khepera robot, which lacks the realistic constraints of a racing vehicle. This study investigates the fundamental informational requirements for proficient racing: is it more effective to utilize absolute spatial coordinates (Newtonian/Third-person) or relative sensory data (Egocentric/First-person)?
Experimental Setup
The researchers developed a 2D simulation environment in Java, modeled after a physical RC racing track at the University of Essex. The track dimensions are $400 \times 300$ pixels.
Five controller architectures were evaluated:
- Action Sequences: A lookup table of 500 discrete actions.
- Open-loop Neural Networks: Inputs limited to the current time step.
- Force Fields: A grid-based system providing speed and orientation vectors.
- Newtonian Neural Networks: Inputs include absolute $(x, y)$ coordinates, velocity, and orientation.
- Sensor-based Neural Networks: Inputs include egocentric data such as speed, wall-distance sensors, and aim-point angles.
Fitness was determined by the number of waypoints (aim points) passed within a 500-step limit. Evolution was tested under both fixed and randomized starting conditions to ensure robustness.
Comparative Analysis and Results
The performance gap between architectures was substantial. The sensor-based (egocentric) controller achieved a peak fitness of 16.04, outperforming both the average human score (11.86) and the best human player in the test group (15.97).
In contrast, the Newtonian neural network failed to develop effective racing lines, often falling into local optima or exploiting flaws in the fitness function. For example, some Newtonian controllers learned to "wiggle" near a waypoint through a wall to accumulate points without actually navigating the circuit.
| Controller | Fixed Start Fitness | Randomized Start Fitness |
|---|---|---|
| Sensor-based (Egocentric) | 13.59 | 12.40 |
| Newtonian Neural Net | 2.94 | 1.84 |
| Open-loop Neural Net | 2.72 | 0.17 |
| Action Sequence | 2.23 | 1.36 |
| Force Field | 1.16 | 0.16 |
Key Insights
The "egocentric advantage" is the study's most significant finding. Mapping absolute third-person coordinates to driving actions is a highly non-linear problem that evolution struggles to solve. Conversely, egocentric data like "distance to the left wall" maps directly to steering adjustments, making it a far more efficient representation for evolutionary algorithms.
Conclusion
Egocentric sensory information is vastly superior for evolving racing controllers. While current models are purely reactive, future research will focus on recurrent neural networks (RNNs) to allow for memory-based racing lines and the use of raw visual input to better simulate human perception.
Summary updated