Stephen Xia

Autoregressive LLM decoding evaluates every generated token through the full layer stack, even though many tokens become predictable at intermediate depths. Existing lossless depth-adaptive methods exploit this redundancy by choosing a single non-final exit depth and verifying its prediction with the final-depth model. However, our measurements show that this selection-based strategy leaves substantial headroom: choosing an exit too late wastes computation, while choosing one too early triggers fallback and discards dependent drafts. We propose Depth Exploration Decoding (DEX), a lossless decoding algorithm that replaces single-depth selection with parallel exploration over multiple candidate depths. At each commit position, DEX validates candidates against the final-depth reference, commits exactly the final-depth token, and collapses the exploration lattice to retain only reusable branch states. This expand--commit--collapse procedure preserves equivalence to standard autoregressive decoding while reducing the cost of committing each token. Across early-exit-trained and standard LLMs, DEX outperforms representative depth-selection baselines and achieves competitive end-to-end throughput against speculative and distributed decoding methods. Moreover, DEX improves as the explored depths become finer, showing that parallel depth exploration provides a scalable way to exploit the underused depth axis of LLM decoding.

Autoregressive LLM decoding evaluates every generated token through the full layer stack, even though many tokens become predictable at intermediate depths. Existing lossless depth-adaptive methods exploit this redundancy by choosing a single non-final exit depth and verifying its prediction with the final-depth model. However, our measurements show that this selection-based strategy leaves substantial headroom: choosing an exit too late wastes computation, while choosing one too early triggers fallback and discards dependent drafts. We propose Depth Exploration Decoding (DEX), a lossless decoding algorithm that replaces single-depth selection with parallel exploration over multiple candidate depths. At each commit position, DEX validates candidates against the final-depth reference, commits exactly the final-depth token, and collapses the exploration lattice to retain only reusable branch states. This expand--commit--collapse procedure preserves equivalence to standard autoregressive decoding while reducing the cost of committing each token. Across early-exit-trained and standard LLMs, DEX outperforms representative depth-selection baselines and achieves competitive end-to-end throughput against speculative and distributed decoding methods. Moreover, DEX improves as the explored depths become finer, showing that parallel depth exploration provides a scalable way to exploit the underused depth axis of LLM decoding.

Speculative decoding accelerates LLM inference by drafting future tokens with a small model, but drafter models degrade sharply under template perturbation and long-context inputs. We identify a previously-unreported phenomenon we call \textbf{attention drift}: as the drafter generates successive tokens within a speculation chain, attention progressively moves from the prompt onto its own recently-generated tokens. We observe this across both \emph{EAGLE3} drafters and \emph{MTP heads}, suggesting drift is a property of drafter designs. We trace this to the un-normalized residual path between chain steps: the drafter's hidden state magnitude grows monotonically with chain depth, which exhibits dynamics consistent with additional pre-norm transformer layers stacked on the target rather than as a standalone autoregressive predictor. In order to limit the growth, we propose two architectural changes: Post-norm on the drafter hidden states and per-hidden-state RMSNorm after capturing target hidden states. Our interventions improve acceptance length over the current leading model, pre-norm EAGLE3, by up to $2\times$ under template perturbation, $1.18\times$ on long-context tasks, and $1.10\times$ on seven standard benchmarks spanning multi-turn chat, math, and coding. Our changes also allow shorter train-time-test depths to generalize over longer drafting sequences.

Speculative decoding accelerates LLM inference by drafting future tokens with a small model, but drafter models degrade sharply under template perturbation and long-context inputs. We identify a previously-unreported phenomenon we call \textbf{attention drift}: as the drafter generates successive tokens within a speculation chain, attention progressively moves from the prompt onto its own recently-generated tokens. We observe this across both \emph{EAGLE3} drafters and \emph{MTP heads}, suggesting drift is a property of drafter designs. We trace this to the un-normalized residual path between chain steps: the drafter's hidden state magnitude grows monotonically with chain depth, which exhibits dynamics consistent with additional pre-norm transformer layers stacked on the target rather than as a standalone autoregressive predictor. In order to limit the growth, we propose two architectural changes: Post-norm on the drafter hidden states and per-hidden-state RMSNorm after capturing target hidden states. Our interventions improve acceptance length over the current leading model, pre-norm EAGLE3, by up to $2\times$ under template perturbation, $1.18\times$ on long-context tasks, and $1.10\times$ on seven standard benchmarks spanning multi-turn chat, math, and coding. Our changes also allow shorter train-time-test depths to generalize over longer drafting sequences.

Achieving multi-lead electrocardiography (ECG) in consumer-grade wearables remains challenging, as most devices provide only a few electrodes and cannot capture spatially diverse cardiac signals. Conventional 12-lead ECG, while clinically standard, requires ten electrodes across the body, confining its use to medical environments. We present EarSleeve, a modular dual-electrode eartip sleeve that transforms off-the-shelf earphones into a 12-lead ECG sensing platform through a human-in-the-loop design. Each sleeve embeds two conductive electrodes and electrically links both sides to form a four-electrode configuration. EarSleeve simultaneously records six limb leads and reconstructs 12-lead–equivalent ECG signals by sequentially contacting standard chest locations. In a 12-user study, EarSleeve captures clear ECG waveforms across all leads and is evaluated against an FDA-cleared reference. Results demonstrate the feasibility of reconstructing 12-lead–equivalent ECG signals using a minimum-electrode configuration under controlled conditions. To our knowledge, EarSleeve is the first system to achieve this with off-the-shelf earphones.

Deploying Large Language Models (LLMs) directly on mobile platforms at the edge is gaining a lot of traction due to a myriad of benefits, such as increased privacy, personalization, and reduced latency. However, LLMs have heavy computational requirements, which are difficult for resource-constrained mobile and edge platforms to fulfill. In addition to limited compute resources, mobile and edge systems often have a compact form factor and lack physical mechanisms to dissipate heat generated from high processor usage rates (e.g., fans) to prevent throttling and reduced processing power, which LLMs can easily cause. To mitigate these effects, prior works have proposed various power governing strategies, such as dynamic voltage and frequency scaling (DVFS), for reducing power and heat generation for heavy computational tasks on mobile platforms. Recently, DVFS methods tailored for mobile LLMs have also been proposed. However, these methods mostly focus on optimizing hardware parameters and processor frequencies, and they fall short under some thermally constrained scenarios. Drawing from recent advances in machine learning, we identify and take advantage of the key insight that not all tokens require full-depth inference to maintain high-quality generation. Motivated by this, we present PELM, a solution that augments traditional DVFS processor frequency tuning with two additional workload-specific knobs: 1) speculative decoding and 2) variable verification depth to expand the optimization space to multiple dimensions for more power efficient on-device LLM inference. In extensive evaluations across different hardware platforms and datasets, PELM demonstrates superior performance compared to state-of-the-art power governing methods, with up to 23.1% speedup and 52.4% reduction in energy consumption, while maintaining comparable task performance. The source code is available at https://github.com/imec-nu/PELM.

Edge devices increasingly run multimodal sensing pipelines that must remain accurate despite fluctuating power budgets and unpredictable sensor dropout. Existing pruning methods fail under these conditions: they generally require fine-tuning after compression, consuming over $10\times$ the deployment energy, and they assign static importance scores that are blind to which sensors are present. We present the SentryFuse framework, which addresses both challenges jointly through two key components. First, SentryGate learns modality-conditioned importance scores during training via first-order saliency supervision and then prunes attention heads and feed-forward channels at deployment without fine-tuning. Second, SentryAttend replaces dense self-attention, a key bottleneck in contemporary multimodal architectures, with sparse grouped-query attention, yielding a net 15% reduction in GFLOPs across three different multimodal architectures. Across three applications and multimodal backbones, SentryGate achieves a 12.7% average accuracy improvement over the strongest pruning baseline, and upto to 18% under modality dropout conditions. Together, SentryFuse reduces memory by 28.2% and lowers latency by up to $1.63\times$ without further fine-tuning, establishing modality-aware zero-shot compression as a practical path to multimodal intelligence on heterogeneous edge hardware.

Edge devices increasingly run multimodal sensing pipelines that must remain accurate despite fluctuating power budgets and unpredictable sensor dropout. Existing pruning methods fail under these conditions: they generally require fine-tuning after compression, consuming over $10\times$ the deployment energy, and they assign static importance scores that are blind to which sensors are present. We present the SentryFuse framework, which addresses both challenges jointly through two key components. First, SentryGate learns modality-conditioned importance scores during training via first-order saliency supervision and then prunes attention heads and feed-forward channels at deployment without fine-tuning. Second, SentryAttend replaces dense self-attention, a key bottleneck in contemporary multimodal architectures, with sparse grouped-query attention, yielding a net 15% reduction in GFLOPs across three different multimodal architectures. Across three applications and multimodal backbones, SentryGate achieves a 12.7% average accuracy improvement over the strongest pruning baseline, and upto to 18% under modality dropout conditions. Together, SentryFuse reduces memory by 28.2% and lowers latency by up to $1.63\times$ without further fine-tuning, establishing modality-aware zero-shot compression as a practical path to multimodal intelligence on heterogeneous edge hardware.

Episodic tasks in Reinforcement Learning (RL) often pose challenges due to sparse reward signals and high-dimensional state spaces, which hinder efficient learning. Additionally, these tasks often feature hidden “trap states”—irreversible failures that prevent task completion but do not provide explicit negative rewards to guide agents away from repeated errors. To address these issues, we propose Time-Weighted Contrastive Reward Learning (TW-CRL), an Inverse Reinforcement Learning (IRL) framework that leverages both successful and failed demonstrations. By incorporating temporal information, TW-CRL learns a dense reward function that identifies critical states associated with success or failure. This approach not only enables agents to avoid trap states but also encourages meaningful exploration beyond simple imitation of expert trajectories. Empirical evaluations on navigation tasks and robotic manipulation benchmarks demonstrate that TW-CRL surpasses state-of-the-art methods, achieving improved efficiency and robustness.

Large Language Models (LLMs) have shown immense human-like capabilities for reasoning and generating digital content. However, their ability to freely sense, interact, and actuate the physical domain remains significantly limited due to three fundamental challenges: (1) physical environments require specialized sensors for different tasks, yet deploying dedicated sensors for each application is impractical; (2) events and objects of interest are often localized to small areas within large spaces, making them difficult to detect with static sensor networks; and (3) foundation models need flexible actuation capabilities to meaningfully interact with the physical world. To bridge this gap, we introduce EmbodiedFly, an embodied LLM agent combining a foundation model pipeline with a reconfigurable drone platform to observe, understand, and interact with the physical world. Our co-design approach features (1) a FM orchestration framework connecting multiple LLMs, VLMs, and an open-set object detection model; (2) a novel image segmentation technique that identifies task-relevant areas; and (3) a custom drone platform that autonomously reconfigures with appropriate sensors and actuators based on commands from the FM orchestration framework. Through real-world deployments, we demonstrate that EmbodiedFly completes diverse physical tasks with up to \(85\%\) higher success rates compared to traditional approaches leveraging static deployments.

Auscultation is a fundamental diagnostic technique, where a physician uses a stethoscope to listen to sounds inside the body for cardiovascular and respiratory assessment. Yet, it requires understanding of what sounds to listen for and where to place the stethoscope, making it inaccessible to the general population. Existing works have attempted to bridge this gap by guiding users on where to place the stethoscope through augmented reality-based technologies. However, these approaches require a camera to understand the current location of the stethoscope and where to move, which raise significant privacy concerns and require adequate lighting conditions. We present an IMU-based alternative that leverages a single inertial measurement unit carried by the user's hand, such as on a stethoscope, to guide users on stethoscope placement and eliminates the need for a camera. Our system tracks the location of the user's hand or stethoscope on the body with less than 8 cm of error, demonstrating the possibility of accurate camera-free guidance.

Accuracy remains a standard metric for evaluating AI systems, but it offers limited insight into how models arrive at their solutions. In this work, we introduce a benchmark based on brainteasers written in long narrative form to probe more deeply into the types of reasoning strategies that models use. Brainteasers are well-suited for this goal because they can be solved with multiple approaches, such as a few-step solution that uses a creative insight or a longer solution that uses more brute force. We investigate large language models (LLMs) across multiple layers of reasoning, focusing not only on correctness but also on the quality and creativity of their solutions. We investigate many aspects of the reasoning process: (1) semantic parsing of the brainteasers into precise mathematical competition style formats; (2) generating solutions from these mathematical forms; (3) self-correcting solutions based on gold solutions; (4) producing step-by-step sketches of solutions; and (5) making use of hints. We find that LLMs are in many cases able to find creative, insightful solutions to brainteasers, suggesting that they capture some of the capacities needed to solve novel problems in creative ways. Nonetheless, there also remain situations where they rely on brute force despite the availability of more efficient, creative solutions, highlighting a potential direction for improvement in the reasoning abilities of LLMs.

Foundation models (FM) have shown immense human-like capabilities for generating digital media. However, foundation models that can freely sense, interact, and actuate the physical domain is far from being realized. This is due to 1) requiring dense deployments of sensors to fully cover and analyze large spaces, while 2) events often being localized to small areas, making it difficult for FMs to pinpoint relevant areas of interest relevant to the current task. We propose FlexiFly, a platform that enables FMs to "zoom in" and analyze relevant areas with higher granularity to better understand the physical environment and carry out tasks. FlexiFly accomplishes by introducing 1) a novel image segmentation technique that aids in identifying relevant locations and 2) a modular and reconfigurable sensing and actuation drone platform that FMs can actuate to "zoom in" with relevant sensors and actuators. We demonstrate through real smart home deployments that FlexiFly enables FMs and LLMs to complete diverse tasks up to 85% more successfully. FlexiFly is critical step towards FMs and LLMs that can naturally interface with the physical world.

We are far from the full adoption of smart home technologies, let alone realizing truly intelligent homes. There are many barriers preventing homeowners from fully adopting smart home technologies, chief among them being their low perceived usefulness (PU) and perceived ease of use (PEoU). Current smart home ecosystems and frameworks only support simple triggers and require homeowners to specify step-by-step logic or purchase new devices to automate new tasks, which severely limits the PU and PEoU of these technologies. We propose DomAIn, a smart home platform that automatically generates, programs, and deploys logic to satisfy a wide range of home-based tasks based on available devices in the home environment, without requiring users to manually "program" any logic. We demonstrate through real deployments and user studies that by incorporating DomAIn, a platform that reduces the need for users to program logic, we can significantly improve a variety of factors that affect PU and PEoU, such as customizability and complexity, by up to 38%, as well as satisfy a diverse range of tasks in home scenarios.

We propose Compressed-Pseudo-Temporal Enhanced Representation Learning (C-PTER), a novel unsupervised clustering framework for human-centered behavior analysis. With the growing prevalence of wearables, smartphones, and IoT devices, vast amounts of human activity data are collected in real-world settings, yet traditional supervised learning approaches require extensive manual labeling, making them impractical for large-scale deployment. Existing deep clustering methods, such as autoencoder-based approaches, often fail to capture temporal dependencies and struggle with noisy sensor readings, leading to suboptimal clustering performance. In contrast, C-PTER integrates pseudo-temporal feature extraction with a parallel CNN-LSTM autoencoder, enabling robust spatial-temporal representation learning. By leveraging compressed feature extraction, our method enhances cluster compactness and inter-cluster separation, significantly improving clustering performance on real-world human activity datasets. We demonstrate that C-PTER outperforms both classical (k-means) and deep clustering baselines (DSC) across three Inertial Measurement Unit (IMU) benchmark datasets (MUser, UCI HAR, MHEALTH), achieving up to 30% improvement in normalized mutual information (NMI) and 21% in accuracy (ACC). These results validate C-PTER as a scalable and effective solution for clustering unsupervised human behavior.

We propose "DUal-NET", a novel transformer-based model for enhancing speech capture through bone conduction headsets in human-centered sensing systems. As wearable bone-conduction devices become increasingly important for continuous health monitoring and ambient computing, they face a unique challenge: bone-conduction microphones can receive significant interference from speakers playing audio to the user, this occurs because the headset is directly in contact with the skull and induces vibrations similar to human speech, much like a user speaking, degrading speech recognition accuracy and communication quality. Existing state-of-art speech enhancement and sound source separation methods are 'blind' and assume that the interference noise is not available due to the inherent difficulty in observing clean correlated noise. By contrast, headsets have full knowledge of the sounds they play through their speakers, and DUal-NET takes advantage of this raw signal in its denoising process. We demonstrate that DUal-NET can significantly improve standard speech quality metrics over existing state-of-art methods in realistic scenarios (PESQ: 135%, STOI: 50%, LSD: 66%), enabling more accurate speech sensing for human-centered applications including health monitoring, personalized assistants, and augmented communication.

Episodic tasks in Reinforcement Learning (RL) often pose challenges due to sparse reward signals and high-dimensional state spaces, which hinder efficient learning. Additionally, these tasks often feature hidden "trap states" -- irreversible failures that prevent task completion but do not provide explicit negative rewards to guide agents away from repeated errors. To address these issues, we propose Time-Weighted Contrastive Reward Learning (TW-CRL), an Inverse Reinforcement Learning (IRL) framework that leverages both successful and failed demonstrations. By incorporating temporal information, TW-CRL learns a dense reward function that identifies critical states associated with success or failure. This approach not only enables agents to avoid trap states but also encourages meaningful exploration beyond simple imitation of expert trajectories. Empirical evaluations on navigation tasks and robotic manipulation benchmarks demonstrate that TW-CRL surpasses state-of-the-art methods, achieving improved efficiency and robustness.

Intimate Partner Infiltration (IPI)--a type of Intimate Partner Violence (IPV) that typically requires physical access to a victim's device--is a pervasive concern around the world, often manifesting through digital surveillance, control, and monitoring. Unlike conventional cyberattacks, IPI perpetrators leverage close proximity and personal knowledge to circumvent standard protections, underscoring the need for targeted interventions. While security clinics and other human-centered approaches effectively tailor solutions for victims, their scalability remains constrained by resource limitations and the need for specialized counseling. We present AID, an Automated IPI Detection system that continuously monitors for unauthorized access and suspicious behaviors on smartphones. AID employs a unified architecture to process multimodal signals stealthily and preserve user privacy. A brief calibration phase upon installation enables AID to adapt to each user's behavioral patterns, achieving high accuracy with minimal false alarms. Our 27-participant user study demonstrates that AID achieves highly accurate detection of non-owner access and fine-grained IPI-related activities, attaining a false positive rate of 1.6%, which is 11x lower than existing methods, and an end-to-end F1 score of 0.981. These findings suggest that AID can serve as a forensic tool that security clinics can deploy to scale their ability to identify IPI tactics and deliver personalized, far-reaching support to survivors.

Speech super-resolution techniques offer a promising solution to enhance audio quality in wearable devices, particularly when addressing the challenges of reduced sampling rates necessitated by battery life constraints and network instability. However, existing methods either prove computationally prohibitive for mobile platforms, or lack sufficient performance. We present TraMSR, a novel hybrid model combining transformer and Mamba architectures for acoustic speech super-resolution. TraMSR achieves superior performance while significantly reducing computational demands compared to state-of-the-art methods. Our model outperforms GAN-based approaches by up to 7.3% in Perceptual Evaluation of Speech Quality (PESQ) and 1.8% in Short-Time Objective Intelligibility (STOI), with an order of magnitude smaller memory footprint and up to 465 times faster inference speed.

Foundation models excel in tasks such as content generation, zero-shot classifications, and reasoning. However, they struggle with sensing, interacting, and actuating in the physical world due to their dependence on limited sensors and actuators in providing timely contextual information or physical interactions. This reliance restricts the system's adaptability and coverage. To address these issues and create an embodied AI with foundation models (FMs), we introduce Embodied Reconfigurable Drone Agent (EmbodiedRDA). EmbodiedRDA features a custom drone platform that can autonomously swap payloads to reconfigure itself with a diverse list of sensors and actuators. We designed FM agents to instruct the drone to equip itself with appropriate physical modules, analyze sensor data, make decisions, and control the drone's actions. This enables the system to perform a variety of tasks in dynamic physical environments, bridging the gap between the digital and physical worlds.

We introduce ElastiFormer, a post-training technique that adapts pretrained Transformer models into an elastic counterpart with variable inference time compute. ElastiFormer introduces small routing modules (as low as .00006% additional trainable parameters) to dynamically selects subsets of network parameters and input tokens to be processed by each layer of the pretrained network in an inputdependent manner. The routing modules are trained using self-distillation losses to minimize the differences between the output of the pretrained-model and their elastic counterparts. As ElastiFormer makes no assumption regarding the modality of the pretrained Transformer model, it can be readily applied to all modalities covering causal language modeling, image modeling as well as visual-language modeling tasks. We show that 20% to 50% compute saving could be achieved for different components of the transformer layer, which could be further reduced by adding very low rank LoRA weights (rank 1) trained via the same distillation objective. Finally, by comparing routing trained on different subsets of ImageNet, we show that ElastiFormer is robust against the training domain.

We propose TRAMBA, a hybrid transformer and Mamba architecture for acoustic and bone conduction speech enhancement, suitable for mobile and wearable platforms. Bone conduction speech enhancement has been impractical to adopt in mobile and wearable platforms for several reasons: (i) data collection is labor-intensive, resulting in scarcity; (ii) there exists a performance gap between state-of-art models with memory footprints of hundreds of MBs and methods better suited for resource-constrained systems. To adapt TRAMBA to vibration-based sensing modalities, we pre-train TRAMBA with audio speech datasets that are widely available. Then, users fine-tune with a small amount of bone conduction data. TRAMBA outperforms state-of-art GANs by up to 7.3% in Perceptual Evaluation of Speech Quality (PESQ) and 1.8% in Short-Time Objective Intelligibility (STOI), with an order of magnitude smaller memory footprint and an inference speed up of up to 465 times. We integrate TRAMBA into real systems and show that TRAMBA (i) improves battery life of wearables by up to 160% by requiring less data sampling and transmission; (ii) generates higher quality voice in noisy environments than over-the-air speech; (iii) requires a memory footprint of less than 20.0 MB.

Automatic drone landing is an important step for achieving fully autonomous drones. Although there are many works that leverage GPS, video, wireless signals, and active acoustic sensing to perform precise landing, autonomous drone landing remains an unsolved challenge for palm-sized microdrones that may not be able to support the high computational requirements of vision, wireless, or active audio sensing. We propose AIRA, a low-cost infrared light-based platform that targets precise and efficient landing of low-resource microdrones. AIRA consists of an infrared light bulb at the landing station along with an energy efficient hardware photodiode (PD) sensing platform at the bottom of the drone. AIRA costs under 83 USD, while achieving comparable performance to existing vision-based methods at a fraction of the energy cost. AIRA requires only three PDs without any complex pattern recognition models to accurately land the drone, under $10$cm of error, from up to $11.1$ meters away, compared to camera-based methods that require recognizing complex markers using high resolution images with a range of only up to $1.2$ meters from the same height. Moreover, we demonstrate that AIRA can accurately guide drones in low light and partial non line of sight scenarios, which are difficult for traditional vision-based approaches.

We propose TRAMBA, a hybrid transformer and Mamba architecture for acoustic and bone conduction speech enhancement, suitable for mobile and wearable platforms. Bone conduction speech enhancement has been impractical to adopt in mobile and wearable platforms for several reasons: (i) data collection is labor-intensive, resulting in scarcity; (ii) there exists a performance gap between state of-art models with memory footprints of hundreds of MBs and methods better suited for resource-constrained systems. To adapt TRAMBA to vibration-based sensing modalities, we pre-train TRAMBA with audio speech datasets that are widely available. Then, users fine-tune with a small amount of bone conduction data. TRAMBA outperforms state-of-art GANs by up to 7.3% in PESQ and 1.8% in STOI, with an order of magnitude smaller memory footprint and an inference speed up of up to 465 times. We integrate TRAMBA into real systems and show that TRAMBA (i) improves battery life of wearables by up to 160% by requiring less data sampling and transmission; (ii) generates higher quality voice in noisy environments than over-the-air speech; (iii) requires a memory footprint of less than 20.0 MB.

Foundation models (FM) have shown immense human-like capabilities for generating digital media. However, foundation models that can freely sense, interact, and actuate the physical domain is far from being realized. This is due to 1) requiring dense deployments of sensors to fully cover and analyze large spaces, while 2) events often being localized to small areas, making it difficult for FMs to pinpoint relevant areas of interest relevant to the current task. We propose FlexiFly, a platform that enables FMs to ``zoom in'' and analyze relevant areas with higher granularity to better understand the physical environment and carry out tasks. FlexiFly accomplishes by introducing 1) a novel image segmentation technique that aids in identifying relevant locations and 2) a modular and reconfigurable sensing and actuation drone platform that FMs can actuate to ``zoom in'' with relevant sensors and actuators. We demonstrate through real smart home deployments that FlexiFly enables FMs and LLMs to complete diverse tasks up to $85\%$ more successfully. FlexiFly is critical step towards FMs and LLMs that can naturally interface with the physical world.

Abstract We introduce a novel human-centric deep reinforcement learning recommender system designed to co-optimize energy consumption, thermal comfort, and air quality in commercial buildings. Existing approaches typically optimize these objectives separately or focus solely on controlling energy-consuming building resources without directly engaging occupants. We develop a deep reinforcement learning architecture based on multitask learning with humans-in-the-loop and demonstrate how it can jointly learn energy savings, comfort, and air quality improvements for different building and occupant actions. In addition to controlling typical building resources (e.g., thermostat setpoint), our system provides real-time actionable recommendations that occupants can take (e.g., move to a new location) to co-optimize energy, comfort, and air quality. Through real deployments across multiple commercial buildings, we show that our multitask deep reinforcement learning recommender system has the potential to reduce energy consumption by up to 8% in energy-focused optimization, improve all objectives by 5–10% in joint optimization, and improve thermal comfort by up to 21% in comfort and air quality-focused optimization compared to existing solutions.

We present the design and implementation of RECA, a novel human-centric recommender system for co-optimizing energy consumption, comfort and air quality in commercial buildings. Existing works generally optimize these objectives separately, or by only controlling energy consuming resources within the building without directly engaging occupants. We develop a deep reinforcement learning architecture based on multitask learning, demonstrate how it can be used to jointly learn energy savings, comfort and air quality improvements for different actions, and build a recommender system with humans-in-the-loop. Through real deployments in multiple commercial buildings, we found that RECA has the potential to further reduce energy consumption by up to in energy-focused optimization, improve all objectives by in joint optimization, and improve thermal comfort by up to in comfort and air quality focused optimization, over existing solutions.

Mapping 3D airflow fields is important for many HVAC, industrial, medical, and home applications. However, current approaches are expensive and time-consuming. We present Anemoi, a sub-$100 drone-based system for autonomously mapping 3D airflow fields in indoor environments. Anemoi leverages the effects of airflow on motor control signals to estimate the magnitude and direction of wind at any given point in space. We introduce an exploration algorithm for selecting optimal waypoints that minimize overall airflow estimation uncertainty. We demonstrate through microbenchmarks and real deployments that Anemoi is able to estimate wind speed and direction with errors up to 0.41 m/s and 25.1° lower than the existing state of the art and map 3D airflow fields with an average RMS error of 0.73 m/s.

Electric vehicles (EVs) have gained widespread popularity in recent years, and the scheduling and routing of EV charging have impacted the welfare of both EV drivers and the grid. In this paper, we present a practical, data-driven, and human-centric EV charging recommendation system at the city-scale based on deep reinforcement learning (DRL). The system co-optimizes the welfare of both the EV drivers and the grid. We augmented and aggregated data from various sources, including public data, location-based data companies, and government authorities, with different formats and time granularities. The data includes EV charger information, grid capacity, EV driving behavior information, and city-scale mobility. We created a 30-day per-minute unified EV charger information dataset with charging prices and grid capacity, as well as an EV driving behavior dataset with location and State of Charge (SoC) information. Our evaluation of the recommendation system shows that it is able to provide recommendations that reduce the average driver-to-charger distance and minimize the number of times chargers switch to a different driver. The dataset we prepared for training the DRL agent, including augmented EV driving data and charging station information, will be open-sourced to benefit future research in the community.

Running with a consistent cadence (number of steps per minute) is important for runners to help reduce risk of injury, improve running form, and enhance overall bio-mechanical efficiency. We introduce CaNRun, a non-contact and acoustic-based system that uses sound captured from a mobile device placed on a treadmill to predict and report running cadence. CaNRun obviates the need for runners to utilize wearable devices or carry a mobile device on their body while running on a treadmill. CaNRun leverages a long short-term memory (LSTM) network to extract steps observed from the microphone to robustly estimate cadence. Through an 8-person study, we demonstrate that CaNRun achieves cadence detection accuracy without calibration for individual users, which is comparable to the accuracy of the Apple Watch despite being non-contact.

The stethoscope is one of the most important diagnostic tools used by healthcare professionals, through a process called auscultation, to screen patients for abnormalities of the heart and lungs. While there are digital stethoscopes on the market which ease this process, it still takes years of training to properly use these devices to listen for abnormal sounds within the body. We present ARSteth, an intelligent stethoscope platform that improves the accessibility of stethoscopes for the general population, allowing anyone to perform auscultation in the comfort of their own homes. Our platform utilizes a combination of augmented reality (AR), acoustic intelligence, and human-machine interaction to dynamically guide users on where to place the stethoscope on different parts of the body (auscultation points), through visual and audio cues. Through user studies, we show that ARSteth, on average, can guide users within 13.2 mm from optimal auscultation points marked by licensed physicians in 13.09 seconds for each auscultation point. By guiding users towards more effective auscultation points, make preventative health screening more accessible and effective for everyone we are able to achieve higher confidence on classifying heart murmurs.

Domain-specific sensor deployments are critical to enabling various IoT applications. Existing solutions for quickly deploying sensing systems require significant amount of work and time, even for experienced engineers. We propose LegoSENSE, a low-cost open-source and modular platform, built on top of the widely popular Raspberry Pi single-board computer, that makes it simple for anyone to rapidly set up and deploy a customized sensing solution for application specific IoT deployments. In addition, the ‘plug and play’ and ‘mix and match’ functionality of LegoSENSE makes the sensor modules reusable, and allows them to be mixed and matched to serve a variety of needs. We show, through a series of user studies, that LegoSENSE enables users without engineering background to deploy a wide range of applications up to 9 × faster than experienced engineers without the use of LegoSENSE. We open-source the hardware and software designs to foster an ever-evolving community, enabling IoT applications for enthusiasts, students, scientists, and researchers across various application domains with or without prior experiences with embedded platforms or coding.