Jonathan Sadeghi

Edward Ayers, Jonathan Sadeghi, John Redford et al.

There is a longstanding interest in capturing the error behaviour of object detectors by finding images where their performance is likely to be unsatisfactory. In real-world applications such as autonomous driving, it is also crucial to characterise potential failures beyond simple requirements of detection performance. For example, a missed detection of a pedestrian close to an ego vehicle will generally require closer inspection than a missed detection of a car in the distance. The problem of predicting such potential failures at test time has largely been overlooked in the literature and conventional approaches based on detection uncertainty fall short in that they are agnostic to such fine-grained characterisation of errors. In this work, we propose to reformulate the problem of finding "hard" images as a query-based hard image retrieval task, where queries are specific definitions of "hardness", and offer a simple and intuitive method that can solve this task for a large family of queries. Our method is entirely post-hoc, does not require ground-truth annotations, is independent of the choice of a detector, and relies on an efficient Monte Carlo estimation that uses a simple stochastic model in place of the ground-truth. We show experimentally that it can be applied successfully to a wide variety of queries for which it can reliably identify hard images for a given detector without any labelled data. We provide results on ranking and classification tasks using the widely used RetinaNet, Faster-RCNN, Mask-RCNN, and Cascade Mask-RCNN object detectors. The code for this project is available at https://github.com/fiveai/hardest.

Jonathan Sadeghi, Nicholas A. Lord, John Redford et al.

Autonomous driving (AD) systems are often built and tested in a modular fashion, where the performance of different modules is measured using task-specific metrics. These metrics should be chosen so as to capture the downstream impact of each module and the performance of the system as a whole. For example, high perception quality should enable prediction and planning to be performed safely. Even though this is true in general, we show here that it is possible to construct planner inputs that score very highly on various perception quality metrics but still lead to planning failures. In an analogy to adversarial attacks on image classifiers, we call such inputs \textbf{adversarial perception errors} and show they can be systematically constructed using a simple boundary-attack algorithm. We demonstrate the effectiveness of this algorithm by finding attacks for two different black-box planners in several urban and highway driving scenarios using the CARLA simulator. Finally, we analyse the properties of these attacks and show that they are isolated in the input space of the planner, and discuss their implications for AD system deployment and testing.

Jonathan Sadeghi, Romain Mueller, John Redford

In engineering design, one often wishes to calculate the probability that the performance of a system is satisfactory under uncertainty. State of the art algorithms exist to solve this problem using active learning with Gaussian process models. However, these algorithms cannot be applied to problems which often occur in the autonomous vehicle domain where the performance of a system may be undefined under certain circumstances. To solve this problem, we introduce a hierarchical model for the system performance, where undefined performance is classified before the performance is regressed. This enables active learning Gaussian process methods to be applied to problems where the performance of the system is sometimes undefined, and we demonstrate the effectiveness of our approach by testing our methodology on synthetic numerical examples for the autonomous driving domain.

Ahmet H. Güzel, Jenny Seidenschwarz, Benjamin Graham et al.

Modern action-conditioned video world models achieve strong short-horizon visual realism, yet remain unreliable on rare, interaction-critical transitions that dominate downstream planning and policy performance. Because passive demonstration data systematically under-samples these high-impact regimes, improving robustness requires actively eliciting model failures rather than relying on their natural occurrence. We introduce a KL-constrained adversarial curriculum in which a policy is trained to expose high-error trajectories of a diffusion-based world model while remaining close to the behavior distribution. The world model is continuously fine-tuned on these adversarially discovered trajectories, yielding an adversarial training loop that converts rare failures into a stable, near-distribution training signal without drifting into out-of-distribution exploitation. To maintain pressure on unresolved weaknesses as the model improves, we propose a Prioritized Adversarial Trajectory (PAT) buffer that re-ranks trajectories based on prediction error, action fidelity, and learning progress, focusing training on unresolved failure modes rather than repeatedly revisiting solved cases. We implement our approach in the MineRL framework and evaluate it on held-out out-of-distribution trajectories; PROWL improves robustness over models trained on passive data alone, reveals reward-hacking behaviors under weak behavioral constraints, and demonstrates that effective adversarial world-model training critically depends on balancing exploratory failure discovery with explicit behavioral regularization. Our results suggest that scalable world models benefit not only from larger datasets, but also from selectively generating informative training data.

Jonathan Sadeghi, Blaine Rogers, James Gunn et al.

There has been increasing interest in characterising the error behaviour of systems which contain deep learning models before deploying them into any safety-critical scenario. However, characterising such behaviour usually requires large-scale testing of the model that can be extremely computationally expensive for complex real-world tasks. For example, tasks involving compute intensive object detectors as one of their components. In this work, we propose an approach that enables efficient large-scale testing using simplified low-fidelity simulators and without the computational cost of executing expensive deep learning models. Our approach relies on designing an efficient surrogate model corresponding to the compute intensive components of the task under test. We demonstrate the efficacy of our methodology by evaluating the performance of an autonomous driving task in the Carla simulator with reduced computational expense by training efficient surrogate models for PIXOR and CenterPoint LiDAR detectors, whilst demonstrating that the accuracy of the simulation is maintained.