Why UDrive

Autonomous Vehicles (AVs) have the potential to change transportation and enhance mobility and road safety while providing environmental benefits through more efficient traffic management1. However, despite the significant technological advancements and investments in recent years, AV companies still face safety and reliability concerns that prevent the large-scale deployment of their fleets. Such challenges can be largely attributed to the inherent issues the traditional Artificial Intelligence (AI) approaches have with uncertainty. Traditional AI models often struggle with making confident predictions when faced with unfamiliar data or real-world scenarios not represented in their training, leading to overconfidence and potential safety risks. UDrive leverages Epistemic AI, a transformative approach that enhances AI systems’ ability to quantify and integrate uncertainty into decision-making. UDrive implements a fully-integrated approach to Perception and End-to-End planning, with both capable of handling uncertainty. The UDrive solution allows AVs to recognise better and respond to unusual objects or challenging conditions, such as poor weather. By comprehensively addressing uncertainty, UDrive promises to enhance safety, robustness, and reliability, leading to better decision-making, improved handling of edge cases, and increased public trust in autonomous driving

How does UDrive work?

AVs typically act with overconfidence when facing unfamiliar or ambiguous situations. Their inability to quantify and communicate epistemic uncertainty poses a major safety risk, as the vehicle may misinterpret critical cues or execute unsafe decisions without recognising its own limitations. Addressing this gap is essential to building AVs that can act like human drivers: aware of uncertainty, cautious in case of ambiguity, and capable of learning from new experiences.

UDrive proposes a ground-breaking solution to effectively handle uncertainty and mitigate its adverse consequences in AVs, ensuring safety and integrity in autonomous driving systems. UDrive leverages the cutting-edge potential of Epistemic AI (E-pi), a new ground-breaking approach that enables AI to be prepared for the data it cannot see, by operating with sets of hypotheses compatible with the (scarce) data available at training time, rather than individual models. This new paradigm is the base for a novel Epistemic E2E planning architecture, which redefines the current
AV software stack based on independent modules, offering a solution for the holistic and continual perception and prediction of road scenes that can be seamlessly integrated within AVs. The new epistemic E2E planning strategy is designed to factor in uncertainties, thus improving the AV system’s robustness and its confidence in making decisions, especially in scenarios involving unknown scene actor behaviour. This approach enables the prediction of semantically-relevant scene elements, including scene object classes, trajectories, and future events, by continually inferring road scene graphs. This continual learning process allows an AV to refine its ability to perceive the world, understand the behaviour of other road agents, and adapt to different environments. Unlike existing perception and planning architectures, UDrive does not just predict; it quantifies its own confidence, enabling AVs to anticipate unknown situations, adapt, and make safer decisions in real-time. This ability to mitigate risk unlocks a unique advantage in safety-critical applications, such as in AVs, supporting market penetration and regulatory approval.

Value Proposition

UDrive positions itself as a disruptive enabler in the autonomous driving market by addressing one of the most critical unresolved challenges: managing uncertainty to ensure safety, robustness, and trust
in autonomous systems. UDrive enables improved object detection, scene classification, and trajectory prediction over SotA systems. Its competitive edge lies in its transparent and interpretable uncertainty models with better domain adaptation and robust tracking of shifts in the data distribution.

UDrive offers:

#1. Intelligent Decision-Making in Unfamiliar Scenarios.

UDrive enables AVs to quantify and act on uncertainty, recognising the limits of their knowledge and making more cautious, context-aware decisions. This leads to greater system resilience, reduced operational risk, and a clear edge in meeting evolving safety and regulatory standards.

#2. Advanced Scene Understanding and Safer Navigation.

UDrive combines uncertainty modelling in both perception and decision-making, enabling more accurate scene interpretation and dynamic risk evaluation in an integrated fashion. This allows AVs to plan safe, human-like manoeuvres in complex environments, with early tests (E-pi project D3.3) showing up to 30% performance gains toward ideal autonomous behaviour.

#3. Continual Learning for Scalable Performance.

UDrive enables AVs to update knowledge while retaining core capabilities, reducing development and data costs, speeding deployment to new markets, and ensuring sustained performance for scalable growth.

#4. Estimation of perception-led risks.

UDrive delivers direct estimation of perception-led risks to vulnerable road users, enabling compliance with safety requirements and giving AV developers auditable safety evidence.

UDrive is an initiative in partnership with Oxford Brookes University and Delft University of Technology. Building on research pioneered by Prof. Fabio Cuzzolin at Oxford Brookes, UDrive advances the development and market readiness of next-generation epistemic AI technologies for autonomous driving