Addressing Uncertainty in Autonomous Driving: Data‑Driven Control Module Strategies

Achieving fully autonomous driving is a complex systems engineering problem that requires precise perception of the environment, understanding the intentions of traffic participants, and stable, safe operation across a wide variety of scenarios. Real‑world road conditions introduce massive uncertainty that spans the entire perception‑prediction‑control pipeline.

The article identifies three major sources of uncertainty:

Perception limitations : limited range of LiDAR, occlusions caused by large vehicles, and hardware constraints.

Behavior prediction randomness : sudden appearances of animals, erratic pedestrians, or other vehicles make future actions highly unpredictable.

Control interaction : when road‑rights are ambiguous, the decisions of different participants create a game‑theoretic interaction that can be subtle or overt.

To give the control module a human‑like “intuition”, the authors propose a data‑driven approach that consists of three layers: problem analysis, algorithm development iteration, and systematic validation.

Problem analysis uses massive, multi‑dimensional data collected from DiDi’s fleet and the “Jushi” cameras (over one million devices, billions of kilometers). Two complementary methods are employed:

Passive analysis : statistical tools identify the most frequent and severe issues, revealing algorithmic and architectural bottlenecks.

Active analysis : indexed queries filter scenes by location, vehicle function, and environmental features, turning interesting cases into training datasets.

Algorithm development relies heavily on machine‑learning, especially deep learning, to process perception outputs. While deep models boost performance, they also become black‑boxes, introducing new sources of uncertainty that conflict with the control module’s stability requirements.

To mitigate this, the authors suggest focusing on well‑defined, data‑rich modules with clear fallback strategies, pursuing continuous learning pipelines for perception and prediction, and exploring explainable AI techniques for long‑term robustness.

Systematic validation must address the non‑exhaustive nature of driving environments. A combination of low‑cost simulation (for rapid early‑stage testing) and real‑world road tests (closed‑track, small‑scale fleet, and open‑city deployments) is required to bridge the gap between simulated and actual performance. The validation workflow includes:

Designing representative scenarios in simulation to detect regressions.

Running closed‑track and on‑road tests to capture edge cases that simulation cannot generate.

Defining a ground truth for planning and control remains an open problem. Human driver data is often used as a proxy, but it may not be sufficient or universally accepted. The authors call for a more rigorous, industry‑wide standard for what constitutes “correct” driving behavior.

In summary, the autonomous driving industry is moving toward hybrid rule‑based and data‑driven strategies to cope with uncertainty. Future progress depends on more advanced data‑driven methodologies, scalable validation pipelines, and collaborative research on explainable, trustworthy AI for control systems.