Caltech CNSE - Zhang - full report

Abstract. A smart car that assists the driver must give warnings in dangerous situations, override the driver to avoid collisions, and help to reach the intended destination as quickly as possible. Unfortunately, satisfying these requirements and at the same time leaving the decisional autonomy at the individual level becomes an extremely hard problem to solve with traditional methods. Biologically-inspired techniques such as Swarm Intelligence and Incremental Evolution provide new promising ways to tackle the design and distributed control problems of a traffic system. In this project, solutions are developed using embodied simulations and validated with real robot experiments.

Motivation. The goal of this project is to work toward developing vehicles endowed with extended sensory capabilities and intelligence that will improve individual and collective traffic safety. The intelligent vehicle will monitor certain road conditions and the driver behavior, give helpful warning signals to the driver in critical situations, and will override the human driver only in emergency situations. In addition to the above constraints, the intelligent monitoring and control system must also be able to cope with a diversity of vehicles and heterogeneous drivers while coping with the noisy and collectively dynamical environment in a robust way.

Research. Satisfying all of these conditions would be a tall order for traditional control algorithms. As a result, we look for inspiration from biological systems. The principal advantage of a biologically inspired approach is that such techniques have stood the test of eons of competition and evolution. Not only are these techniques robust, they also have the advantage of scalable and distributed operation, as well as acceptance of existing heterogeneous agents. A specific biologically inspired approach that seems well optimized for understanding collective phenomena (like automobile traffic) is Swarm Intelligence. Swarm Intelligence provides a framework in which autonomy, emergence, and distributed robustness replace centralized control. This is analogous to comparing flocking birds to a complex man-made air-traffic control system that results in countless flight delays and lost luggage. To implement the principle of Swarm Intelligence, an evolutionary approach is used here to explore the traffic situations and optimize the control system in simulation.

Achievements. With support from Delphi Automotive Systems, sample traffic situations are simulated in the Webots 2.0 robot simulator, which is a 3D embodied, sensor-based simulator. A simplified model of a human driver, which is aware of its own speed, position, orientation, and what lies within its field of view, will try to avoid other cars and decide either to follow or change lanes. If, for whatever reason, the simulated human driver causes the car to enter any undesirable dangerous situation, the driver will first be warned by the system. Only when the situation becomes dire and requires immediate evasive action, will the complementing controller override the human driver. In all other cases, the commands given by the driver (steering wheel angle, gas/brake pedal deflection, etc) are passed directly to the actuators. The controller will get data from on-board distance sensors and lane sensors in order to have an awareness of the state of the environment. Incremental evolutionary techniques are used to suggest the optimal placements and configurations for the sensors on the vehicle as well as other warning and control parameters. The initial simulations are conducted on a straight three-lane highway (see Figure 1) but curved streets may be added later.

Figure 1. Screenshot of Webots 2.0 simulation program

Currently, the Webots simulator does not simulate the holonomicity of real automobiles. Specifically, Webots 2.0 was designed to simulate small, circular, unicycle robots. But the real traffic situation can be scaled to the simulation. The future version of Webots (3.0 or higher) simulator will allow the users to construct the simulated robots in their own way and incorporate the real vehicle dynamics, hence enabling a more realistic simulation. To study the traffic fluidity in very large macroscopic simulations, cellular automata-based platforms (e.g. TRANSIMS) may eventually be leveraged. Finally, implementing this system in real robots (see Figure 2) would provide concrete confirmation or refutation of any results obtained from simulation.