The transition to 5G and 6G networks has led to a widespread adoption of machine learning (ML) for critical functions like modulation classification, channel estimation, resource management, and spectrum sensing. While ML has enhanced operational efficiency, it has simultaneously expanded the attack surface for adversarial ML at the Physical Layer (PHY), for example, from Generative Adversarial Networks (GANs). While techniques like radio frequency (RF) fingerprinting have emerged as a PHY-level authentication method based on hardware-induced signal traits (such as in-phase/quadrature (I/Q) imbalance and error vector magnitude), GANs can synthesize RF signals to mimic legitimate hardware-induced features up to 95% similarity. This is close enough to evade most detection schemes. Existing defenses to GANs based on convolutional neural networks, deep neural networks, supervised retraining, and/or heuristics do not generalize well across different modulations, protocols, channel conditions, or unseen attack types. Autoencoder and reconstruction-based approaches are often limited to clean reference signals, which are not always available in dynamic wireless environments. While GANs are excellent at mimicking low-order statistics (mean/variance), they fail to replicate complex signal structures.

Machine learning has transitioned from traditional supervised learning to more resource-efficient sketching and federated techniques. Early compressive learning relied on hand-crafted random projections and task-specific iterative solvers. While these methods reduced data volume, they were inflexible because a change in data distribution or task required a complete redesign of the projection. Concurrently, privacy-preserving needs led to the rise of federated learning and differential privacy. However, these methods often struggled with high communication costs and the inability to merge model updates effectively across different architectures. Until recently, the state of the art remained fairly bifurcated, where one could have either high-accuracy iterative training on raw data, or efficient but brittle and task-specific compressed representations that lacked generalizability across diverse analytical tasks, e.g., Principal Component Analysis (PCA), regression, clustering.

An autonomous, solar-powered robot designed to travel along suspended wires for long-term, non-invasive environmental monitoring in hard-to-reach natural areas.

Achieving seamless robotic interaction with physical environments requires a sophisticated blend of sensory perception and logical reasoning. UC Berkeley researchers have developed "RealWorldPlay," a physical artificial intelligence system designed to enhance robotic action through a unified multimodal reasoning framework. The system integrates a visuo-tactile policy—combining sight and touch—with a large language model (LLM) that provides real-time verification feedback and strategic planning. By utilizing a "world model" to generate self-training data, the platform allows robots to autonomously set goals and learn from simulated scenarios, ensuring that their physical actions are both reasoned and verified before execution.

Brief description not available

The Global Navigation Satellite System (GNSS) consists of a family of satellite navigation systems (like GPS, Galileo, GLONASS, BeiDou) which provide global positioning and navigation from orbiting satellites. GNSS is one of the major inputs for phone location. Accurate pedestrian localization in “urban canyons” has long been limited by GNSS multipath errors and blocked line-of-sight, especially for blind and low-vision pedestrians who need sidewalk-level accuracy. GNSS-based positioning in dense downtowns is often limited to tens of meters off because skyscrapers block satellites, create multipath, and reduce signal quality, leading to especially large errors that make it hard to know which side of a street a pedestrian is on. For blind and low‑vision users, conventional smartphone navigation (pure GNSS, camera‑based visual positioning system, or beacon infrastructure) does not offer reliable, hands‑free, street‑side-accurate guidance. Most accuracy-focused approaches to date require detailed 3D models, specialized hardware, and/or substantial map annotations, limiting the scalability across urban environments and challenging mainstream apps deployment. Moreover, for blind and low‑vision pedestrians, integrating precise localization with usable, low‑attention interaction (i.e., no constant camera use, minimal screen looks) and robust crossing guidance is still a problem.

Monitoring the structural integrity of buildings traditionally requires expensive, specialized sensor networks that are difficult to deploy at scale. UC Berkeley researchers have developed a novel approach that leverages the existing network of smartphones equipped with the MyShake earthquake early warning application. By utilizing the highly sensitive accelerometers within millions of consumer devices, the system measures the natural frequencies and damping ratios of buildings through ambient vibrations. This crowdsourced data provides a real-time, large-scale assessment of structural health across entire urban environments. The platform effectively transforms everyday mobile devices into a distributed seismic monitoring array, allowing for continuous observation of building performance without the need for dedicated hardware installations.

SHIELD leverages a customized large language model pipeline to detect and investigate sophisticated cyber threats with high accuracy and interpretability.

An intent-aware, dual-stream AI architecture that adapts compute allocation and inference depth on embedded platforms, balancing rapid triage and detailed analysis for real-time visual understanding.

Researchers at the University of California, Davis have developed a method to design optimized current profiles for lithium-ion batteries using analytic sensitivity functions. By leveraging a reduced electrochemical model, the approach enables fast and accurate identification of key parameters, improving battery management systems and reducing testing time.

Researchers at the University of California, Davis have developed a navigation system that constructs a sensing map from wireless signal observations and pedestrian deadreckoning (PDR) data to enable accurate indoor navigation without relying on traditional geographic localization maps.

A multimodal, wireless wearable device enabling continuous and detailed stress assessment and subclassification.

Traditional stool samples provide an indirect and often "blurred" snapshot of the complex microbial environment within the human gut, making it difficult to design precise health interventions. UC Berkeley researchers have developed InferBiome, a computational framework that reconstructs the actual state of the gut microbiome from stool data. By inverting a blurring model and applying a probability-based simulation of microbiome dynamics, the system predicts how different dietary interventions will impact an individual's unique gut ecosystem. This method allows for the selection of personalized dietary recommendations that maximize host health benefits by simulating outcomes across various possible microbiome states.

Brief description not available

Brief description not available

Brief description not available

Brief description not available

Brief description not available

Brief description not available

Brief description not available

AI infrastructure that collects, transcribes, and analyzes student audio responses to deliver actionable insights on learning experiences.

Brief description not available

UC Berkeley researchers have developed a sophisticated computer-implemented framework that leverages transformer architectures to model the evolution of biological sequences over time. Unlike traditional phylogenetic models that often assume sites evolve independently, this framework utilizes a coupled encoder-decoder transformer to parameterize the conditional probability of a target sequence given multiple unaligned sequences. By capturing complex interactions and dependencies across different sites within a protein or genomic sequence, the model estimates the transition likelihood for each position. This estimation allows for a high-fidelity simulation of evolutionary trajectories. This approach enables a deeper understanding of how proteins change across different timescales and environmental pressures.

Researchers at the University of California Davis have developed an unlearning method that precisely removes specific data influences from trained large language models while preserving their overall knowledge and performance.

Historically, speech therapy apps have relied primarily on online cloud-based speech recognition systems like those used in digital assistants (Cortana, Siri, Google Assistant), which are designed to best guess speech rather than critically evaluate articulation errors. For children with cleft palate specifically, affecting 1 in 700 babies globally, speech therapy is essential follow-up care after reconstructive surgery. Approximately 25% of children with clefts use compensatory articulation errors, and when these patterns become habituated during ages 3-5, they become particularly resistant to change in therapy. Traditional approaches to mobile speech therapy apps have included storybook-style narratives that proved expensive with low replayability and engagement, as well as fast-paced arcade-style games that failed to maintain user interest. Common speech therapy applications require a facilitator to evaluate speech performance and typically depend on continuous internet connectivity, creating barriers for users in areas with poor network coverage or those concerned about data privacy and roaming costs. The shift toward gamified therapy solutions showed that game elements can serve as powerful motivators for otherwise tedious activities. Speech recognition systems face inherent limitations in accuracy compared to cloud-based solutions and require substantial processing power and memory that can impact device performance and battery life, particularly on older mobile devices. Automatic speech recognition (ASR) models struggle significantly with children's speech due to non-fluent pronunciation and variability in speech patterns, with phoneme error rates reaching almost 12%, and consonant recognition errors affecting the reliability of speech disorder detection. The challenge becomes even more pronounced for populations with speech impairments, as conventional ASR systems are optimized for typical adult speech rather than atypical articulation patterns of cleft palate speech or developmental disabilities. Moreover, maintaining user engagement over extended therapy periods is hard, and many apps fail to provide sufficient motivation for daily practice, which is essential for speech improvement.