September 27, 2023
 min read

Navigating the Autonomous Driving Landscape: Technical Challenges and the Path Forward with Themis AI

How Capsa helps to address technical challenges in autonomous driving.

 Navigating the Autonomous Driving Landscape: Technical Challenges and the Path Forward with Themis AI

In this blog post, we look at the business case for autonomous vehicles and we show how Capsa can solve some of the challenges faced by this industry. 

Autonomous vehicles are key factors in shaping the future of transportation and related markets. Below are some numerical data to better see the market influence of this technology:

  • Approximately 65% of the nation's consumable products are transported to markets by trucks. Research indicates that the implementation of full autonomous vehicles would lead to a reduction in operating expenses by around 45%, resulting in cost savings ranging from $85 billion to $125 billion for the US trucking sector. Similar predictions show that the entire logistic industry will be massively impacted by these changes.
  • Robo-taxis and autonomous shuttles promise to lower the cost of transportation in the future, as they will likely be more affordable than individually-owned vehicles. After introducing taxis in some parts of the country, robo-taxis companies now have the goal to scale by rapidly deploying robo-taxi fleets across multiple cities in the US. The value of the Robotaxi Market is projected to reach $ 45.7 billion by 2030.
  • Projections suggest that by 2035, the autonomous driving market will reach an economic impact between $300 to $400 billion. A recent McKinsey Insights analysis indicates that for consumers there are several advantages in switching to an autonomous single-car market: it could revolutionize the way consumers experience mobility (allowing them to spend more time for other, more rewarding activities rather than driving); it would increase safety reducing the number of accidents; it has the potential to reduce energy consumption and costs, thanks to the autonomous system's capability to consistently maintain optimal speeds.

Adoption Bottlenecks and Trust Paradigms in Autonomous Driving Tech 

Despite early adopter enthusiasm, future adoption trajectory estimates do not look as rosy as expected. Here some evidence of this decreasing excitement around autonomous vehicles adoption. According to a recent S&P Global Mobility report, adoption of autonomous vehicles is slower in pace than expected and it will mostly concern vehicles with “conditional automation” (Level 3), aka vehicles where the system manages steering and acceleration, monitors roads and driving surroundings, but needs a human driver to take over and execute specific real-time operations and maneuvers.

Furthermore, research shows a decreasing trend in trust metrics, with a decline from 35% in 2020 to 26% in 2021 of consumers considering autonomous vehicles transition. Some events have also strengthened consumers’ skepticism and lack of trust. In 2023, the California Public Utilities Commission issued permits and allowed two companies to operate fleets of autonomous robo-taxis around the clock in San Francisco and collect passenger fares for these rides. This is a technology with a level 4 autonomy, where vehicles can autonomously navigate a predefined area without the need for human monitoring and manual control. Almost immediately, some of these vehicles started to behave unexpectedly and dangerously, putting people and other vehicles in jeopardy. In particular, one autonomous vehicle was involved in a collision with a fire truck, prompting outrage and further scrutiny regarding the challenges of seamlessly integrating self-driving cars into cities and urban settings.

More broadly, we believe that there are three main hurdles autonomous vehicles need to tackle before they can become mainstream and invert some lowering trends: 

  • Computational Intensiveness: autonomous vehicles necessitate computationally intensive algorithms for perception, prediction, and planning. This computational load is expensive and also produces carbon emissions analogous to the totality of existing data centers, as shown in a recent MIT study.
  • Trust in Algorithmic Infrastructure: the downward trend in adoption consideration underscores the imperative of establishing robust trust paradigms in autonomous systems. One route to adopt to solve this problem is to accelerate the establishment of a standard, government-approved framework for evaluating the safety of autonomous vehicles.
  • Algorithmic Consistency and Robustness: ensuring robust operational performance across a myriad of driving environments and conditions remains a pressing challenge for scaling both the development and adoption of autonomous vehicles.

Capsa: A Technical Vanguard in Autonomous Navigation

Themis AI, with its cutting-edge technology, offers a plausible solution to the technical and safety-related challenges of autonomous driving. Our proprietary algorithms, honed over five years of rigorous full-scale vehicle experimentation, have achieved:

• A 16-fold decrement in collision incidents.

• A twelvefold acceleration in algorithmic computation timelines.

• A recovery rate of 89% in near-crash scenarios.

• A 93% reduction in manual human intervention instances.

Research in the field of autonomous driving has shown deep neural networks (DNN) can lead to high accuracy in autonomous vehicles functionalities. Unfortunately, accuracy is possible as long as there is an ample supply of training data whereas failure is to be expected in out distribution domains. When implemented in real-world scenarios, control systems for autonomous vehicles have to face high levels of unpredictability and unfamiliar environments.

Many datasets for training and testing consist mainly of cars driving down straight roads in great weather conditions with high visibility. In some cases, however, vehicles will face conditions of bumpy roads, adverse weather, and low visibility with near-collision scenarios. These are the situations in which the model is most likely to fail. Unfortunately, these conditions are usually scarcely represented in training sets. 

Therefore, a world where autonomous vehicles are safely and reliably integrated requires that those systems are aware of the domains in which they lack knowledge, and are able to provide a correct confidence level for their outputs. We call these ‘vacuitic’ and ‘epistemic’ uncertainties, and Themis AI’s Capsa provides a suite of tools to measure these uncertainties. As the research behind our technology shows, a system able to provide an estimation of ‘vacuitic’ and ‘epistemic’ uncertainties, is able to detect “novel events which the network has been insufficiently trained for and not trusted to produce reliable outputs.” By measuring uncertainty, Capsa can determine a model's robustness, namely its ability to perform reliably and accurately in diverse scenarios. In addition, this technology will lead to “automated debiasing of a neural network training pipeline, leading to faster training convergence and increased accuracy.” That means that Capsa empowers organizations and AI developers to take a significant step toward the ultimate goal of ensuring that autonomous technology can be truly trusted. 

In conclusion, by strategically integrating their uncertainty estimation algorithms with state-of-the-art technology, Themis AI has elevated the operational proficiency of autonomous vehicles and expanded their applicability spectrum. As we work towards enhancing computational efficiency, fostering trust, and refining algorithmic robustness, we can get closer to the ultimate goal of an autonomous vehicular ecosystem.

 Navigating the Autonomous Driving Landscape: Technical Challenges and the Path Forward with Themis AI

Latest articles

View all

Don’t keep up with AI, stay ahead of it.

Work with us