As self-driving, autonomous, vehicles head out on to public roads, one of the field’s most persistent challenges remains collision avoidance in unpredictable traffic. A study in the International Journal of Vehicle Design discusses an artificial intelligence (AI) control system that has a 97 per cent success rate in avoiding obstacles, with a maximum response time of about half a second.
Urban roads present a shifting landscape of pedestrians, stalled vehicles, roadworks and erratic drivers. For a self-driving car, safe operation depends not only on accurate sensors but also on rapid decisions made under such uncertain conditions. Conventional obstacle-avoidance systems often rely on fixed rules or straightforward processing of sensor data. These approaches can sometimes fail in heavy rain, fog, or headlight glare.
Other systems that use reinforcement learning, a branch of AI in which the algorithm learns by trial and error, such as Deep Deterministic Policy Gradient, need a lot of computing power and often struggle to work quickly enough for real-world driving conditions.
The new approach described in IJVD builds on a reinforcement learning framework called Soft Actor-Critic, or SAC. In this computing system, the software actor proposes driving actions while the software critic evaluates whether or not the given manoeuvre would be sensible or not. SAC is designed to learn so that positive outcomes boost the actor-critic interactions that led to them. The system also embeds entropy, a statistical measure of randomness that allows it to continue to explore the best manoeuvres rather than settling prematurely on a single solution. This helps the system remain adaptable in uncertain environments.
The model also incorporates a self-organising cluster mechanism inspired by the collective movement of a flock of birds, that famously avoid mid-air collisions. At close range, a mathematically defined repulsion force pushes vehicles apart to prevent impact. At medium distances, a velocity calibration rule aligns speed with an ideal braking curve to reduce the risk of rear-end collisions. Additional rules govern wall and obstacle avoidance. This layered design allows multiple autonomous vehicles to coordinate their movements without relying on a single lead vehicle.
Ma, Y., Qian, Y., Ma, T., Li, Y. and Wan, J. (2025) ‘Intelligent obstacle avoidance control method for autonomous vehicles based on improved SAC algorithm’, Int. J. Vehicle Design, Vol. 99, No. 5, pp.1–19.
No comments:
Post a Comment