Precise evaluation of vehicle CO₂ emissions under real-world driving conditions is essential for meeting stringent regulations. While laboratory procedures such as WLTC ensure repeatability, they fail to reflect transient conditions typical of everyday driving, and PEMS-based RDE measurements remain costly. This study proposes a data-driven approach for reconstructing instantaneous CO₂ concentration using onboard engine parameters - engine speed, exhaust gas temperature and exhaust mass flow rate - without direct CO₂ sensing. Linear, nonlinear and ensemble machine learning models were evaluated using an RDE dataset of 5200 synchronized observations collected on a mixed urban-rural route. Ensemble methods, particularly Random Forest (R² = 0.715, RMSE = 15,307 ppm) and XGBoost (R² = 0.669), achieved the highest accuracy and reproduced steady-state conditions and rapid CO₂ drops during fuel-cut events. The results confirm that reliable CO₂ estimation can be achieved using a minimal OBD-based input set, enabling cost-effective emission monitoring and real-time onboard applications.