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.
We present a live demonstration of the SPHERE CPS enclave, a reconfigurable industrial-control experimentation environment built to make ICS security studies repeatable and shareable across research teams. The enclave combines PLC-based control, modular I/O, SCADA/HMI interaction, and an isolated, configurable network segment, with optional integration of simulated components to emulate larger process contexts. This demo walks attendees through an end-to-end workflow using a water-treatment scenario: provisioning the control and monitoring stack, running a baseline process, introducing representative faults or adversarial manipulations, and collecting synchronized controller, network, and process telemetry for replay and offline analysis. The goal is to support evaluations of detection and response techniques that depend on coupled cyber and physical behavior, including controller-logic attacks and network-mediated disruptions. We conclude by outlining planned extensions toward richer operator-in-the-loop studies (e.g., mixed-reality interfaces) and broader cross-domain interactions spanning water and energy-related infrastructure.
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.
