Learning predictive models from high-dimensional sensory observations is fundamental for cyber-physical systems, yet the latent representations learned by standard world models lack physical interpretability. This limits their reliability, generalizability, and applicability to safety-critical tasks. We introduce Physically Interpretable World Models (PIWM), a framework that aligns latent representations with real-world physical quantities and constrains their evolution through partially known physical dynamics. Physical interpretability in PIWM is defined by two complementary properties: (i) the learned latent state corresponds to meaningful physical variables, and (ii) its temporal evolution follows physically consistent dynamics. To achieve this without requiring ground-truth physical annotations, PIWM employs weak distribution-based supervision that captures state uncertainty naturally arising from real-world sensing pipelines. The architecture integrates a VQ-based visual encoder, a transformer-based physical encoder, and a learnable dynamics model grounded in known physical equations. Across three case studies (Cart Pole, Lunar Lander, and Donkey Car), PIWM achieves accurate long-horizon prediction, recovers true system parameters, and significantly improves physical grounding over purely data-driven models. These results demonstrate the feasibility and advantages of learning physically interpretable world models directly from images under weak supervision.
We introduce a novel framework for learning context-aware runtime monitors for AI-based control ensembles. Machine-learning (ML) controllers are increasingly deployed in (autonomous) cyber-physical systems because of their ability to solve complex decision-making tasks. However, their accuracy can degrade sharply in unfamiliar environments, creating significant safety concerns. Traditional ensemble methods aim to improve robustness by averaging or voting across multiple controllers, yet this often dilutes the specialized strengths that individual controllers exhibit in different operating contexts. We argue that, rather than blending controller outputs, a monitoring framework should identify and exploit these contextual strengths. In this paper, we reformulate the design of safe AI-based control ensembles as a contextual monitoring problem. A monitor continuously observes the system's context and selects the controller best suited to the current conditions. To achieve this, we cast monitor learning as a contextual learning task and draw on techniques from contextual multi-armed bandits. Our approach comes with two key benefits: (1) theoretical safety guarantees during controller selection, and (2) improved utilization of controller diversity. We validate our framework in two simulated autonomous driving scenarios, demonstrating significant improvements in both safety and performance compared to non-contextual baselines.
Cyber-physical systems (CPS) increasingly face security threats that can disrupt critical infrastructure operations. The SPHERE CPS enclave is a modular, remotely accessible industrial control system (ICS) testbed designed to support security experimentation on programmable logic controllers (PLCs), industrial networks, and digital twin simulations. It enables researchers to investigate cyber-physical attacks, anomaly detection, and intrusion resilience strategies. Unlike general cybersecurity testbeds, SPHERE's CPS enclave provides a configurable, realistic environment for studying adversarial scenarios that bridge cyber and physical domains. The infrastructure offers controlled, reproducible experiments with customizable network topologies and hardware-in-the-loop validation. This poster presents the design philosophy, community-driven experimental goals, and deployment considerations of the SPHERE CPS enclave, demonstrating its potential for advancing CPS security research.
In wearable optical sensing systems, non-idealities present during measurement affect the quality of acquired signals of interest. The issue is of particular significance in deep tissue sensing applications in which the signal of interest is faint relative to typical noise sources that contaminate the measured data. To mitigate the problem, we investigate the role of optimizing the light source activation pulse, specifically its toggling rate and duty cycle. We identify the potential for the optical source activation signal to mitigate the problem and characterize its design space. Subsequently, we present JOIN, an algorithm that yields the parameters of an optimized activation signal for a given profile of contextual noise profile. Through this algorithm, we strike a balance between avoiding noisy spectral regions associated with environmental disturbances and fitting a judiciously-selected number of signal harmonics into the measurement bandwidth, thereby enhancing the signal-to-noise ratio (SNR) of the acquired data from deep tissue by as much as 6dB. The proposed method is validated through analytical derivations, as well as test bench and in-vivo measurements on human subjects.
Driving is a safety-critical activity, with most accidents caused by human errors. Advances in artificial intelligence (AI) offer the potential to reduce these errors, yet AI systems still face challenges in making socially acceptable decisions under complex and dynamic circumstances. To overcome these challenges, incorporating ethic-by-design principles is essential in developing autonomous systems. A crucial component of ethic-by-design is liability determination, which remains difficult to automate due to the need for retrospective analysis of vague and subjective criteria. In this paper, we propose a model-based liability determination framework that integrates legal doctrines with a driver behavior model. Our framework first formalizes liability determination based on established legal principles. The driver behavior model enables retrospective analysis by identifying reasonable alternative actions at each decision point, facilitating the assessment of duty breaches and proximate causes. We validate our framework through experiments focusing on highway driving accidents. We selected 8 representative simulated accidents and evaluated them with 5 senior traffic polices. The results demonstrate that our framework's liability judgments closely align with those of the police, indicating high accuracy. Additionally, we examined the impact of breach of duty judgments on liability determination by utilizing Large Language Models (LLMs). The findings reveal that with only trajectories, LLM's liability determinations were inaccurate, whereas incorporating these judgments resulted in outputs that matched police assessments. Our proposed liability determination framework effectively automates the process, providing accurate and interpretable results that support ethic-by-design in autonomous driving systems, thereby enhancing their safety and accountability.
This paper presents a unique sim-to-real assessment environment for traffic signal control (TSC), LibSignal++, featuring a 14-ft by 14-ft scaled-down physical replica of a real-world urban roadway equipped with realistic traffic sensors such as cameras, and actual traffic signal controllers. Besides, it is supported by a precise indoor positioning system to track the actual trajectories of vehicles. To generate various plausible physical conditions that are difficult to replicate with computer simulations, this system supports automatic sensor manipulation to mimic observation changes and also supports manual adjustment of physical traffic network settings to reflect the influence of dynamic changes on vehicle behaviors. This system will enable the assessment of traffic policies that are otherwise extremely difficult to simulate or infeasible for full-scale physical tests, providing a reproducible and low-cost environment for sim-to-real transfer research on traffic signal control problems.
Connected vehicle applications like platooning, supported by Vehicle-to-Everything (V2X) communication, help improve road safety and traffic efficiency through coordination. Departing from IEEE 802.11p-based Dedicated Short Range Communication (DSRC) standard for V2X, 6G-enabled Joint Communication and Sensing (JCAS) with cellular V2X (cV2X) promises to meet the strict latency requirements of connected vehicles by optimizing communication based on sensed vehicle distribution. This work proposes a novel framework for co-simulating vehicle dynamics control with V2X communication enabled by JCAS. The framework enables RL-driven dynamic adaption of the communication beam based on vehicle position and density, leading to better V2X communication and related vehicle dynamics.
Verification of Cyberphysical Systems (CPS) is fundamental to design and deployment of devices, and it relies on specification-based monitoring through expert engineers devised properties, subsequently expressed through several forms of formal logic (e.g., first order, propositional, temporal) depending on the needed semantics. In fact, any monitoring-based approach is only effective if the formal property accurately reflects the requirements originally elicited (mostly in natural language) by testing engineers who may lack familiarity with logic and formal methods. An intelligent translation process is needed to connect property engineering with formal monitoring and verification approaches. In this poster, we propose a gray box Natural Language to Temporal Logic translation module. It conditions large language models with temporal logic examples and provides an intermediate representation along with the formal translation of the natural language property. This intermediate representation, a simplified abstract syntax tree, encodes relationships between components of the requirement. We establish quantifiable criteria for example quality in the training and testing sets, including metrics for formality and specification length. Finally, we demonstrate how such structural information consistently leads to improved accuracy and show the impact of requirement complexity on the proposed approach's performance against state-of-the-art LLM-based benchmarks.
In smart Internet of Things (IoT) applications, edge devices often collect and store limited data, which is insufficient for training modern deep learning models. Collaborative training methods like cloud computing and federated learning enable robust models for IoT applications, yet introduce data privacy concerns due to central data collection and model inversion attacks. Remedies such as differential privacy can bring data privacy protection but dramatically degrade the accuracy performance of IoT applications. To safeguard user data privacy while maintaining application quality, it is imperative to establish a framework capable of preserving user privacy without compromising accuracy standards. In this paper, we present Atlas, a private and accurate personalized federated local differential privacy (LDP) framework for IoT applications. We first design a layer-sharing mechanism called the layer importance mask to separate the local model into global and personalized layers. Second, we design a weighted LDP mechanism and add noise to the global layers before transmitting them to the federated learning framework for aggregation. Third, we combine local personalized layers and aggregated global layers to perform IoT tasks. Our experiments on five real-world IoT application datasets and the CIFAR-10 dataset showthat our privacy-preserving approach only sacrifices 2% to 6% of accuracy compared to the state-of-the-art non-privacy preserving FL frameworks among various IoT applications and outperforms the current LDP-based FL framework by 8% to 13%.
Predicting spatiotemporal patterns is essential for traffic flow forecasting in Intelligent Transportation Systems (ITS), where accurate predictions can greatly enhance traffic management. Currently, data-driven approaches, such as Graph Neural Networks (GNNs) combined with physics-informed partial differential equations (PDEs), have shown promising performance in capturing complex traffic dynamics. However, these models sometimes make unexpected and incorrect predictions with high certainty which will mislead real-world decision-making, particularly in critical scenarios. This lack of uncertainty quantification (UQ), which estimates the confidence of neural network predictions beyond prediction accuracy, limits the reliability of the predictions. Existing UQ methods in traffic forecasting are typically applied to purely data-driven models, leaving the effects of physics-informed approaches on UQ largely unexplored. In this paper, we address these challenges by combining multiple UQ baselines with physics-informed methods to investigate the impact of physical constraints on UQ for traffic forecasting. Additionally, we introduce a new physics-informed loss function to guide the model learning process, which holds an exponential relation and complements the linear relations of the PDE layer. Through extensive experiments on real-world traffic datasets, we demonstrate that our proposed method outperforms existing approaches, achieving reductions of up to 14.8% in short-term and 8.7% in long-term traffic speed prediction errors. Sensitivity analysis further illustrates the robustness of our approach under perturbations.
An e-taxi fleet consumes a significant amount of energy daily, making it a substantial electricity consumer. Unlike traditional consumers, such as factories and buildings, a fleet coordinates charging activities across both times and locations, offering considerable flexibility in its energy demand. This allows a fleet to achieve substantial reductions in energy consumption in response to demand response requests while maintaining transportation service quality. To better understand and control this intrinsic energy flexibility, we propose the eFlx framework for managing e-taxi fleets for demand response. In the eFlx framework, we establish a model to characterize the energy flexibility upon receiving a real-time demand response request. We then investigate the energy flexibility provisioning problem, formulated as a bi-level optimal control problem, which aims to optimize and maintain the energy flexibility of the fleet for potential demand response requests that could arise at any time. To achieve real-time flexibility provisioning, we develop an efficient iterative algorithm to solve this problem. Data-driven evaluations with NYC datasets demonstrate that eFlx achieves a 19. 98% greater reduction in energy demand compared to existing solutions, without requiring extra charging or compromising the quality of taxi service.
The goal of this paper is to estimate whether a human or Adaptive Cruise Control (ACC) is managing a vehicle's speed control, based on observations by external sensors. The driving characteristics of individual vehicles-whether human-driven or ACC-controlled-play a crucial role in shaping overall traffic flow. To enable advanced traffic control strategies tailored to specific vehicle behaviors, this paper introduces a time-series deep learning classifier that leverages multiple models, including One-Dimensional Convolutional Neural Networks (1D-CNN), Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Temporal Fusion Transformers (TFT). These models distinguish between human-driven and ACC-controlled trajectories using signals such as the ego vehicle's velocity, the distance to the leading vehicle, and derived features. Unlike previous studies relying solely on simulation data, our classifier uses large-scale, real-world datasets from field experiments and daily commute data. By utilizing low-latency, low-anomaly signals decoded from Controller Area Network (CAN) bus messages, the model achieves a high accuracy of 98.85% in classifying human-driven and ACC-controlled vehicles within three seconds, outperforming existing methods that require longer trajectory data or pre-calibrated models. The approach is scalable and can be integrated with large-scale traffic trajectory datasets, such as those from the I-24 Motion project, enabling more precise estimation of ACC penetration, fuel consumption, and emissions.
In fields such as autonomous and safety-critical systems, online optimization plays a crucial role in control and decision-making processes, often requiring the integration of continuous and discrete variables. These tasks are frequently modeled as mixed-integer programming (MIP) problems, where feedback data are incorporated as parameters. However, solving MIPs within strict time constraints is challenging due to their $\mathcal{NP}$-complete nature. A promising solution to this challenge involves leveraging the largely invariant structure of these problems to perform most computations offline, thus enabling efficient online solving even on platforms with limited hardware capabilities. In this paper we present a novel implementation of this strategy that uses counterexample-guided inductive synthesis to split the MIP solution process into two stages. In the offline phase, we construct a mapping that provides feasible assignments for binary variables based on parameter values within a specified range. In the online phase, we solve the remaining continuous part of the problem by fixing the binary variables to the values predicted by this mapping. Our numerical evaluation demonstrates the efficiency and solution quality of this approach compared to standard mixed-integer solvers, highlighting its potential for real-time applications in resource-constrained environments.
We present algorithms for Cyber-Physical Systems (CPS) falsification and control, which take advantage of knowing the entire language of the temporal logic specification - that is, the set of signals that satisfy the formula. In the design of CPS, falsification and control play key roles. Falsification is a testing task, where the goal is to find an input signal that causes the system's output trajectory to violate the correctness requirements. Control is the dual task, where the goal is to find an input signal that causes the system's output to satisfy the specification. When the specification is expressed in a temporal logic, most existing work relies on local optimization heuristics to perform both tasks. In this paper, we explore whether a different expression of the specification offers advantages when performing falsification and control. Recent work presented a method for computing a representation of the language of a formula in (discrete-time) Signal Temporal Logic (STL), showing that the language can be represented as a union of polytopes. We introduce new falsification algorithms which combine distance information to the different components of the language to accelerate the convergence to a falsifier. And we introduce a new algorithm for computing a satisfying control signal which works by repeatedly projecting violating output trajectories back onto the language's components. Moreover, these algorithms are trivially parallelizable to take advantage of multiple processors. Despite their relative simplicity, our algorithms demonstrate 10x to 100x speedups relative to the state-of-the-art.
We present a cost-effective prototype rover for farms built on a ride-on toy car platform. Equipped with cost-effective components and operable via multiple user interfaces, our prototype rover offers farms of various scales an affordable entry into precision agriculture.
Adversarial methods like false data injection attacks (FDIAs) on sensors can manipulate system states in modern cyber-physical systems (CPSs). Such attacks may remain undetected by state-of-the-art statistical detectors, leading to an unsafe system state. In this work, we propose a novel detector neural network with the ability to differentiate between distributional shifts of sensor data under normal operation and adversarial attacks. This includes stealthy attacks that can bypass state-of-the-art detectors. The proposed scheme helps in securing resource-constrained safety-critical CPSs using lightweight solutions that employ compute and communication-heavy crypto methods for data exchange only upon detection of false data injection by an adversary.
Neural network (NN) control has been adopted widely in cyber-physical systems (CPS). When an NN-based policy fails a formally specified task, NN repair algorithms can fix it. Recent literature raises the problem of Repair with Preservation (RwP), which requires preserving existing correct behaviors while repairing the incorrect ones; a corresponding solution is given, known as Incremental Simulated Annealing Repair (ISAR). In this paper, we tackle the computational efficiency issue of ISAR, which involves expensive log-barriered objective functions and wastes computational efforts rolling back when a repaired NN breaks correct behaviors. With our analysis, we reduce the RwP problem to a stability-plasticity (S-P) trade-off interpolation problem, which has been studied in continual learning (CL). Then, we propose our method, ISAR with Interpolation (ISAR-I), which majorly improves ISAR. ISAR-I abandons the expensive log barriers and rolls back to allow intermediate policies to compromise correct behaviors for repair. Then, an interpolation of the S-P trade-off between the original NN and the intermediate NN is kicked off in the Bayesian space, searching for a final NN that both repairs and preserves. Case studies in OpenAI Gym mountain car and an unmanned underwater vehicle show that ISAR-I is able to preserve all verified trajectories while repairing 81.7% and 21.3% of the broken ones, respectively, achieving the same performance as ISAR, with runtime cost of only 6.5% and 19.6%, on average. Source code: https://github.com/ericlupy/isar_interpolation
Haptic bilateral teleoperation holds promise for applications such as telemaintenance, remote manipulation, and disaster response, yet delivering precise, low-latency force and video feedback remains challenging. This study advances haptic bilateral teleoperation by combining live video with Model Mediated Teleoperation (MMT) to enable predictive force feedback. While this method has benefits, several non-trivial challenges, such as synchronizing the model with user's and remote robot's actions, arise. A novel algorithm is developed that allows the robotic device to replicate interactions predictively experienced by the operator. We validated this approach in a fully functional system that performs reliably despite significant network delays. The latency performance of the system is extensively characterized, achieving a motion-to-pixel latency of 58 ms. A user study revealed that operators did not perceive network latency of at least 75 ms, resulting in a 133 ms motion-to-pixel delay requirement. Additionally, a 5G latency analysis demonstrated that effective haptic teleoperation is achievable with both operator and remote ends connected via 5G. This provides a path away from strict latency requirements toward practical teleoperation solutions using currently available technology.
