Publications

Mapless UAV navigation in complex environments: A survey on deep reinforcement learning approaches
Dawei Ni
Steven H. H. Ding
Benjamin C. M. Fung
As applications of Unmanned Aerial Vehicles (UAVs) continue to expand across sectors such as logistics, surveillance, and environmental moni… (voir plus)toring, the development of safe and robust autonomous navigation capabilities has become increasingly important. Mapless autonomous UAV navigation refers to a UAV’s ability to reach a designated destination along a collision-free and efficient path without human intervention and without relying on a global map. Achieving this capability is particularly challenging in complex environments, especially in cluttered spaces, in the presence of dynamic obstacles, or under wind disturbances. Deep reinforcement learning (DRL) has emerged as one of the leading approaches for modeling UAVs as intelligent agents that learn effective control policies through interaction with the environment. However, the performance of these methods depends strongly on both the characteristics of the environment and the limitations of the underlying algorithms. A deep understanding of algorithmic trade-offs is essential for selecting and tailoring the appropriate RL framework to ensure safe, robust and efficient UAV navigation in complex environments. Within this context, this study provides a structured overview of recent advances in state-of-the-art DRL approaches for Mapless UAV navigation and reviews their application across various UAV-environment dynamics and constraints. In addition, the review identifies key challenges in deploying DRL in highly complex environments. It also outlines open research questions and proposes future research directions to advance this field.
A model-based adaptive sampling framework for infectious disease surveillance
Supun Manathunga
Nathaniel Osgood
Sheila F O’Brien
David L Buckeridge
W. Alton Russell
Longitudinal surveillance must choose each round's sampling design before the next observations arrive. Fixed sample-size rules treat each r… (voir plus)ound as an isolated cross-sectional survey and ignore the evolving uncertainty in the epidemic process, which changes as transmission, vaccination, interventions, variants, and behavior change. We present a model-based adaptive sampling framework for choosing the next-round design. A state-space model represents the epidemic process and maps latent states to surveillance estimands. A statistical filter calibrates the model as observations arrive, propagating uncertainty in latent states, time-varying parameters, stochastic dynamics, and sampling noise. A Bayesian assurance criterion converts the one-step-ahead predictive distribution into the minimum-cost design likely to achieve a target posterior credible-interval width. We evaluated the framework in two simulations and in a retrospective case study of SARS-CoV-2 serosurveillance among Canadian blood donors in four regions, replaying the series with each week's specimens subsampled to the selected design. In simulations, the policy increased sampling effort during an anticipated high-variance period, and the share of intervals meeting the width target (0.970) agreed with validated assurance (0.974). After an unanticipated change in transmission rate, validated assurance stayed at 0.960 while the share meeting the target fell to 0.894 (95% Monte Carlo interval 0.869 to 0.919) in the transition window. Sampling effort rose while the filter relearned the transmission rate, then fell. In the case study, recommended sample sizes tracked the one-step predictive uncertainty and were largest during the Omicron wave, and the 90% filtering interval had width at most 0.10 in 94% of region-rounds with specimens. The modular framework extends to other pathogens, data streams, and surveillance objectives.
AIFaultBench: A Reproducible Benchmark of Real-World Faults in AI Software Systems
Mehil B Shah
Mohammad Masudur Rahman
AI-based software systems remain difficult to debug because their failures can involve interactions among code, models, data, dependencies, … (voir plus)hardware, and execution environments. Existing benchmarks cover specific classes of AI systems, while concerns about fault realism and reproducibility highlight the need for documented execution conditions and evidence. We present AIFaultBench, a benchmark of 770 real-world bugs from 105 AI-based software repositories across 76 organizations. It covers six classes of AI systems: machine learning, deep learning, reinforcement learning, systems based on large language models, agentic AI, and AI tooling and infrastructure. Each bug record connects the original report and affected codebase to structured metadata, environment setup scripts, a reproduction trigger, and investigation evidence. We reproduced 652 bugs (84.7%) under the recorded execution conditions and documented the outcomes of the remaining 118 attempts. Through its broad coverage, documented reproduction evidence, and common structure, AIFaultBench supports empirical studies of AI faults, debugging tool evaluation, practical failure investigation, and teaching with real-world examples.
Chance constrained optimization of hub transport systems
Max Bourgeat
Öykü Naz Attila
Rémy Chevrier
Antoine Legrain
Railway transportation offers an efficient and reliable service between stations. However, in remote areas, reduced accessibility due to lar… (voir plus)ge distances between stations and passengers’ origin/destination points presents a significant challenge for railway companies. In low-density areas and in the countryside in particular, the supply of complementary transport to the train is often limited, or even nonexistent during off-peak hours, and does not always allow for an intermodal journey. For the French railway operator (SNCF), it is desirable to be able to provide a complementary offer to transport service in order to make a typical passenger journey as seamless as possible. A key aspect to operate this service successfully is to ensure that shuttles arrive at the train station before the time that passengers are due to take their train. Consequently, finding an optimal routing plan for shuttles is a highly time-sensitive problem. Besides, an inaccurate representation of travel times can easily lead to infeasible routing decisions with respect to meeting the due date that passengers are required to be at the train station. Given this context, the main contribution of this paper is the formalization of this problem, that we refer to as the hub transportation problem with chance constrained due dates . The problem involves optimizing routing decisions for shuttles while ensuring their timely arrival at the train station under uncertain travel times. The second contribution is an algorithm based on column generation to solve it. The pricing problem incorporates chance constraints, deeming a route feasible if the shuttle can reach the train station before the passengers’ due date. Travel time uncertainty is modeled using a discrete and finite uncertainty set to represent various travel time scenarios. The chance constraints ensures that the final decision remains feasible by meeting the passengers’ due date in at least a specified proportion of the travel time scenarios. Because solving the chance-constrained pricing problem exactly on a large scenario set is intractable, the proposed heuristic algorithm samples a finite scenario set, accelerates the labelling algorithm in the pricing problem with reduced dominance and feasibility sets, and recovers an integer solution by re-solving the master problem with the generated columns. Experiments on 100 instances built from real transit and census data, with synthetic travel-time scenarios generated from a calibrated multimodal distribution, show that the chance-constrained approach consistently meets the desired empirical feasibility rates while achieving cost savings compared to deterministic counterparts, which tend to be either overly conservative or insufficiently reliable.
RepGen: Reproducing Deep Learning Bugs through Context-Aware Code Generation and Verification
Mehil B Shah
Mohammad Masudur Rahman
Reproducing deep learning (DL) bugs is challenging because failures can involve source code, models, data, dependencies, and execution envir… (voir plus)onments. Incomplete bug reports further complicate reproduction by omitting essential details. We present RepGen, an open-source tool for reproducing DL bugs through context-aware code generation and iterative verification. Given a bug report and repository, RepGen explores the codebase to build a learning-enhanced context, develop a reproduction plan, and generate candidate scripts. It refines candidates using static checks and LLM feedback to improve their correctness and relevance. When enabled, the tool prepares the execution environment, runs the script, and assesses its output against the reported failure. RepGen provides an interactive command-line interface supporting local and hosted LLMs, configurable execution budgets, and inspectable artifacts, along with a GitHub Action for automated issue processing. The underlying approach reproduced 85 of 106 real-world DL bugs (80.19%). A controlled study with 27 developers showed a 23.35% improvement in reproduction success and a 56.8% reduction in reproduction time.
On the Limits of Test-Time Compute for Self-Evaluation
Self-verification and aggregation have emerged as popular approaches for bridging the gap between pass@1 and pass@N. However, increasing the… (voir plus) number of candidates does not consistently narrow the gap between final-answer accuracy and pass@N. To understand this limitation, we use ranking error to measure how reliably models distinguish their correct generations from incorrect ones. We show that ranking error admits a bias-variance decomposition, with both components shaped by algorithmic design choices and inference compute. Across our experiments, additional compute reduces variance, while bias initially decreases before approaching a plateau. Guided by this analysis, we introduce Scorpio, a simple and general method that jointly scores randomly grouped candidates. Across math, code, and puzzle benchmarks, Scorpio improves verification, tree search, and aggregation. For verification, Scorpio improves accuracy over the strongest baseline by up to 7 percentage points while matching or exceeding that baseline's peak accuracy with 17-42% lower inference cost. For tree search, using Scorpio to score partial solutions and select the final answer improves accuracy by up to 10 percentage points over prior methods. Finally, Scorpio improves the scaling of aggregation-based methods by 11 percentage points on coding tasks.
Algorithmic exclusions, collective inclusions: surveying algorithmic harms and collective action in LGBTQIA2S+ communities
Mina Alfaghih
Elliot Creager
Annie Pullen Sansfaçon
Pixelated Posterior Sampling of Point-spread Functions in Astronomical Images
Abstract We introduce a novel framework for upsampled point-spread function (PSF) modeling using pixel-level Bayesian inference. Accurate PS… (voir plus)F characterization is critical for precision measurements in many fields, including weak lensing, astrometry, and photometry. Our method defines the posterior distribution of the pixelized PSF model through the combination of an analytic Gaussian likelihood and a highly expressive generative diffusion model prior, trained on a library of Hubble Space Telescope effective PSF (ePSF) templates. Compared to traditional methods (parametric Moffat, ePSF template-based, and regularized likelihood), we demonstrate that our PSF models achieve orders-of-magnitude-higher likelihood and residuals consistent with noise, all while remaining visually realistic. Further, the method applies even for faint and heavily masked point sources, merely producing a broader posterior. By recovering a realistic, pixel-level posterior distribution, our technique enables the first meaningful propagation of detailed PSF morphological uncertainty in downstream analysis. An implementation of our posterior sampling procedure is available on GitHub.
The global geography of threats to freshwater biodiversity
Christophe Brabant
Jonas Geldmann
Mette Bendixen
Lise Comte
Harith M. Farooq
Katherine Hébert
Bernhard Lehner
Catherine Overed-Sayer
Lars Lønsmann Iversen
Human activities are eroding biodiversity, especially in freshwater systems. Despite major advances in quantifying species vulnerability and… (voir plus) threats, we still lack a crucial understanding of where threats are affecting freshwater species globally. Here, we produced a species-centered spatial assessment of threats, mapping the impact of 45 threats across 21,490 freshwater species of amphibians, fish, mammals, odonates, and reptiles. We found that species in highly developed and water-managed regions are strongly affected by multiple overlapping threats. Although threats vary across taxa and regions, increasing threat probabilities elevate extinction risk and population declines, with freshwater species experiencing greater threat intensity and distinct spatial patterns than terrestrial species. Our findings establish a spatially explicit baseline for targeting the most urgently needed freshwater conservation and management interventions.
Competitive EV charging station location with queues
The Minh Nguyen
Amira Dems
Electric vehicle (EV) public charging infrastructure planning faces significant challenges in competitive markets, where multiple service pr… (voir plus)oviders affect congestion and user behaviour. This work extends existing modelling frameworks by incorporating the presence of competitors’ stations and more realistic queueing systems. First, we analyse three finite queueing systems, M/M/1/K,M/M/s/K, and M/Er/s/K, with varying numbers of servers (charging outlets) and service time distributions, deriving analytic expressions for user behaviour metrics. Second, we embed the queueing-based user behaviour model into a bilevel program, where the upper level locates new charging stations to maximise accessibility (throughput), and the lower level captures users’ station choices via a user equilibrium. Third, we apply a reformulation from competitive congested user-choice facility location models to approximately solve the bilevel problem and introduce a surrogate-based heuristic to enhance scalability. Fourth, we showcase our methodology on a real-world case study of an urban area in Montreal (Canada), offering managerial insights into how user-choice behaviour assumptions and competition affect throughput and location decisions. The results demonstrate that our model yields (re)location strategies that outperform the existing network. More broadly, this approach provides a tool for incorporating charging service quality—through queueing metrics—and existing competition into station planning.
GradSplat: Gradients Inside the Forward Pass for Feed-Forward 3D Gaussian Splatting
Jingong Chen
Guangyuan Li
Xinxin Zuo
Chulwoo Pack
Kyle Gao
Kwanghee Won
Feed-forward 3D Gaussian Splatting predicts a scene from a few images, but our predictor is less accurate than its own Gaussians optimized p… (voir plus)er scene on those images. This gap lies mostly in the shape of the Gaussians, their scale and rotation, since optimizing the shape alone recovers almost all of it and opacity and color recover less. Gradient steps on the input views correct this shape, and the model holds these views at inference. Log-anisotropy correction changes the axis ratios of each Gaussian in log space at constant volume, and in-forward gradient refinement renders the Gaussians into the input views and applies the gradient of the photometric loss with a learned recurrent update, first to opacity and color and then to the shape. GradSplat is a feed-forward model that refines its own Gaussians by learned gradient steps inside its forward pass, with no target view. On both complete six-view RealEstate10K protocols, it has the best PSNR and SSIM among nine methods and keeps its accuracy from
Semantic Anchor Transport: Robust Test-Time Adaptation for Vision-Language Models
Shambhavi Mishra
Julio Silva-Rodríguez
Ismail Ben Ayed
Jose Dolz
Large pre-trained vision-language models (VLMs) like CLIP exhibit strong zero-shot performance but struggle under distributional shifts. We … (voir plus)propose Semantic Anchor Transport (SAT), a method that generates pseudo-labels for test samples by aligning visual embeddings with reliable text-based semantic anchors using Optimal Transport for batch-wise label assignment. These pseudo-labels enable efficient test-time adaptation through principled cross-modal alignment. We further incorporate multi-template distillation to leverage diverse textual clues, replicating multi-view contrastive learning without added computational cost. Extensive experiments demonstrate consistent performance gains over state-of-the-art methods across multiple benchmarks while maintaining computational efficiency.