The race to self-driving cars: Why Waabi World says its ‘AI-first’ approach could win

If any company finally makes autonomous vehicles mainstream, it will make billions. But in humans, learning to drive requires an elusive mix of instinct, muscle memory, and quick reactions – and transferring these skills to machines is no easy task. 

The road to self-driving cars is full of hazards, like safety issues, legal concerns, and privacy violations. 

But Canada-based startup Waabi, founded by University of Toronto professor Raquel Urtasun, says its “AI-first” driving simulator “Waabi World” has the edge over existing training methods. 

Why training autonomous vehicles is such a challenge 

First of all, self-driving AI requires a particularly large amount of data and power. 

The deep neural networks used in self-driving vehicles demand immense computational resources. Training driverless models requires vast datasets covering countless factors, including weather conditions, possible hazards, and other driving scenarios. 

Accuracy is particularly crucial in self-driving AI systems. These models must achieve detection and prediction accuracy beyond those expected in typical AI applications, which means rigorous testing and validation. 

Handling unpredictable situations poses a particular challenge. Consider atypical situations like a traffic officer waving drivers through a red light, sudden and extreme changes in weather, or horses traveling on a busy road. 

As such, rule-based programming can fall short when trying to “teach” an autonomous vehicle how to react to real-world dangers. There are simply too many variables. 

Common approaches to training autonomous vehicles 

Let’s look at some approaches commonly used to train autonomous vehicles. Most existing self-driving technologies draw upon one or more of these techniques. 

Sensor fusion 

Sensor fusion involves combining data from various sensors (such as radar, lidar, and cameras) to create a comprehensive view of the road and driving conditions. 

Waymo (formerly Google’s self-driving project) uses sensor fusion extensively in its autonomous vehicles. 

Because sensor fusion draws on so many data sources, it can create a holistic impression of what’s happening on the road. Even if one sensor fails in a given situation, another can compensate. This approach is particularly advantageous in low-visibility weather conditions, such as fog or heavy rain. 

However, sensor fusion is costly, and aligning so many types of data can be challenging. 

Deep learning 

Deep learning algorithms train autonomous vehicles using vast amounts of data to make real-time decisions.  

NVIDIA develops AI-based self-driving platforms that leverage deep learning and neural networks. 

The continuous ingestion of and learning from data means that a model can continuously adapt and self-improve. 

However, high-quality training data is crucial. Deep learning models will struggle to adapt to unknown situations. And the system operates as a “black box”—it can be impossible to understand the rationale for good or bad decisions. 

HD maps and localization 

High-definition maps provide precise road information, and localization techniques help ensure accurate positioning within the environment. 

HERE Technologies specializes in mapping and location services, including HD maps for self-driving cars. 

These maps can be extremely detailed and can keep autonomous vehicle models up-to-date with information about lane markings and road signs. But conversely, keeping HD maps updated can be a challenge. A lot of infrastructure is required to map factors such as roadworks and traffic conditions. 

V2X Communication 

Vehicle-to-everything (V2X) communication involves vehicles exchanging data with infrastructure and other vehicles.  

Audi has been actively researching and implementing V2X communication in its vehicles. 

V2X technology helps provide early warnings about traffic conditions, road closures, and hazards. But there are privacy and security concerns—many drivers might be uncomfortable with their car transmitting a continuous stream of their journey. 

Waabi’s autonomous vehicle training method 

Waabi World seeks to solve many of the challenges presented by the above methods through a combination of deep learning, probabilistic inference, and complex optimization. 

Unlike traditional methods that result in a software stack requiring manual tuning, Waabi says its approach is end-to-end trainable, interpretable, and capable of complex reasoning. 

Testing autonomous vehicles on real roads can be costly, dangerous, and present legal difficulties. Waabi World is an advanced simulator that serves as a virtual training ground for self-driving software. By using simulation, Waabi reduces the need for real-world testing. 

The ‘scale’ problem 

Waabi aims to solve the problem of scale in self-driving car research. Traditional methods struggle with less frequent and unpredictable driving scenarios. By their nature, such scenarios occur rarely, meaning self-driving cars have little opportunity to respond and adapt to them. 

Waabi claims that its approach streamlines self-driving development because it allows unpredictable scenarios to present more frequently than they otherwise would. In theory, this approach should be more cost-effective and technically feasible. 

An ‘AI-first’ approach 

Waabi calls its approach “AI-first”, which likely means a greater focus on machine learning and less focus on sensory data. 

As such, the company has employed several prestigious technologists, such as ex-Google computer scientist Geoffrey Hinton and ImageNet developer Fei-Fei Li. 

In the race to crack self-driving cars, an AI-first approach could be hugely efficient and beneficial – if Waabi pulls it off.