Autonomous Driving is Hard: Safe decision-making in uncertain environments

In 1966, Stanford Research Institute (SRI International) invented Shakey, the first robot to make decisions about its own actions.  Fast-forward to now, news of self-driving cars making poor decisions, leading to harm to individuals, bringing cities to a halt, and impeding emergency services is making headlines for the wrong reasons. Why is decision-making  so difficult for these systems?
To answer that, we need to understand the kinds of decisions required in driving. As humans, when we drive, our priority is always to reach our destination safely and as efficiently as possible. While we may tolerate delays, we rarely compromise on safety. The same condition applies to autonomous vehicles. Driving decisions can be categorized into three levels: strategic, tactical, and operational. Each level serves a distinct purpose and operates on different time-frames: 

Strategic decisions focus on high-level and long-term (minutes to hours) goals like choosing the best route to a destination. 

Tactical decisions deal with mid-term (seconds to minutes) context-dependent adjustments to adapt to the current environment while following the strategic plan, such as merging onto a motorway or overtaking a slower vehicle. 

Operational decisions are rather short-term (milliseconds to seconds) and responsible for immediate control and execution of tactical decisions like pressing the brake pedal to stop at a red light or steering slightly to stay centered in the lane. For autonomous vehicles to succeed, they must excel at all three levels, especially the tactical layer that bridges high-level planning with real-time execution.

However, tactical decision-making remains a significant hurdle. The complexity of tactical decision-making becomes especially apparent when designing algorithms capable of human-like judgment in nuanced, high-pressure scenarios, such as navigating complex roundabouts. It remains one of the most pressing challenges in the journey toward full autonomy. 

Tactical decision-making in autonomous vehicles is a highly complex process, heavily dependent on the vehicle's perception systems. For example, merging onto a motorway requires the vehicle to understand both its own state (e.g. velocity, position, etc.) and the states of surrounding vehicles. Using this information, it must then determine whether merging is safe. Unlike humans, however, autonomous vehicles perceive their environment through sensors like cameras. An onboard computer then processes and translates this data into representations of objects. Due to external factors like poor lighting or internal factors like sensor errors, the vehicle might therefore misinterpret what it "sees". It could, for example, confuse a pedestrian for a cyclist, mistake someone on the road for someone on the sidewalk, or fail to identify objects altogether. These perception errors can significantly hinder effective tactical decision-making and compromise safety.

Also, tactical decision-making is the core of situational adaptability. Situational adaptability requires handling a massive variety of scenarios in real time, such as deciding whether to yield, overtake, or merge. This involves interpreting dynamic and uncertain contexts like traffic flow, road conditions and interactions with other road users. These decisions need to balance safety, performance, and comfort with clear, deterministic outcomes. For example, overtaking a slow-moving vehicle improves performance but might temporarily reduce safety margins or the decision to turn right at a junction with oncoming vehicles must consider the trade-off between waiting longer (performance) and making the turn (safety risk).

Having access to accurate, reliable and relevant information is crucial for autonomous decision-making. If an autonomous vehicle had complete and perfectly accurate knowledge of its surroundings, making safe and reliable decisions would be much easier. However, in the real world, this is rarely the case. Factors like extreme weather conditions or sensor limitations can lead to incomplete or inaccurate data. Relying directly on such information without accounting for its impact on decision-making can result in unsafe choices, potentially jeopardizing safety. Unlike skilled human drivers, who instinctively infer missing details using common sense and experience—such as anticipating a hidden pedestrian behind a stopped bus or assuming a vehicle in the next lane might cut in. Autonomous vehicles, however, lack these intuitive abilities therefore it’s important to explore how they can still make safe decisions under uncertainty.

Tactical decision-making lies at the core of achieving safe, reliable, and human-like autonomous driving. Without robust tactical capabilities, strategic planning loses relevance, and operational control falls short. Tackling the challenges of tactical decision-making requires advancing algorithms that can handle uncertainty, predict outcomes, and adapt to changing conditions. While humans excel at making complex decisions with minimal effort, replicating this ability in machines demands extensive research and innovation.

Autonomous vehicles may already be on our roads, but the journey to match human-level decision-making is far from over. In the downloadable example below I highlight how uncertainty and time constraints transform even simple scenarios into complex decision problems. To enable safe and effective decision-making in dynamic environments, significant progress is needed. By focusing on the tactical layer, we address the heart of the challenge, bridging the gap between autonomy’s promise and its real-world application.

Download my example scenario: Autonomous Driving decision making - a real-world scenario (PDF , 323kb)