Prediction of time-to-collision for autonomous ground vehicles: An evaluation of supervised and reinforcement learning approaches

Time-to-collision (TTC) prediction is an essential factor for autonomous ground vehicle (AGV) safety frameworks and decision-making. This research focuses on establishing an efficient benchmark evaluation of supervised machine learning, Neural Networks (NN) and deep reinforcement learning approaches with diverse operational datasets. The investigation utilises a diverse feature set and validation protocol to evaluate three major machine learning (ML) families, namely supervised learning, neural network-based learning, and reinforcement learning. The categorisation includes linear ML regressors (such as Ordinary Least Squares (OLS), lasso, ridge, and ElasticNet), composite ensemble regressors (gradient boosting, random forest, and AdaBoost), Neural network-based (multi-layer perceptron, deep neural network), a meta-learner (MLP-Hybrid) and deep RL algorithms (Deep Q Network, Deep Deterministic Policy Gradient). Key performance indicators (regression scores) indicate that reinforcement learning models significantly outperform classical approaches. The Deep Deterministic Policy Gradient (DDPG) achieved the highest predictive accuracy ( R ²  = 0.986), followed by MLP-Hybrid and Deep Q Network (DQN) with R ² of 0.891 and 0.878, respectively. Neural network models yield an average of 0.857, while linear and ensemble models delivered average results of 0.752 and 0.727, respectively. The results indicate that harnessing continuous-action reinforcement delivers higher precision for the prediction of TTC and offers robustness for handling complex driving scenarios.