Archan Misra

Photoplethysmography (PPG) is widely used in non-invasive health monitoring applications such as heart rate and blood pressure estimation. Despite deep learning substantially advancing PPG sensing accuracy, models trained on a population often struggle with cross-user generalization , exhibiting significant performance degradation when applied to unseen individuals. Building on the evidence that PPG signals encode biometric traits for user authentication, we hypothesize that these identity-specific features are a primary confounding noise of poor cross-user generalization. To validate this hypothesis and address the issue, we propose PPG-IDR (Identity Disentangled Representations), a framework designed to disentangle medical physiological features from identity-specific information. PPG-IDR utilizes a dual-branch design to partition the feature space, employing adversarial and orthogonality constraints to suppress identity leakage, alongside a self-supervised objective to refine medical representations. We evaluated PPG-IDR using multiple datasets across six downstream tasks, and the results demonstrate that PPG-IDR consistently outperforms strong baselines in unseen-user scenarios. These findings highlight the importance of identity disentanglement for scalable and robust cross-user PPG sensing.

Photoplethysmography (PPG) has become a cornerstone of physiological sensing in wearable devices, enabling non-invasive monitoring of heart rate and related biomarkers. However, its reliability deteriorates sharply under dynamic, high-intensity, or non-periodic motions such as those in sports, where existing datasets fail to capture realistic wrist dynamics. To address this gap, we introduce PPG-Sport, the first large-scale dataset designed for heart rate monitoring from wrist-worn PPG under real sports conditions. The PPG-Sport dataset includes synchronized PPG, inertial measurement unit (IMU), and electrocardiography (ECG) recordings from both wrists of 30 participants across six representative activities: stationary, walking, running, badminton, table tennis, and basketball, amounting to 48 hours of multimodal data. PPG-Sport uniquely captures three critical properties absent in prior datasets: (1) non-periodicity, reflecting irregular and broadband motion patterns; (2) high intensity, with frequent, large-magnitude accelerations; and (3) bilateral asymmetry, caused by distinct functional roles of the dominant and non-dominant hands during sports. We further establish a deep-learning-based benchmark that combines both temporal and spectral representations of PPG and IMU signals to evaluate heart rate estimation performance. Experimental results show that models trained only on conventional periodic activities fail drastically in sports scenarios. Although incorporating sports data mitigates the degradation, significant errors remain. These findings highlight PPG-Sport as an essential and challenging benchmark for developing motion-robust physiological sensing algorithms during sports activities. Dataset and benchmark code are available at: https://github.com/LaserHu/PPG-Sport.

Embodied agents struggle to generalize to new environments, even when those environments share similar underlying structures to their training settings. Most current approaches to generating these training environments follow an open-loop paradigm, without considering the agent's current performance. While procedural generation methods can produce diverse scenes, diversity without feedback from the agent is inefficient. The generated environments may be trivially easy, providing limited learning signal. To address this, we present a proof-of-concept for closed-loop environment generation that adapts difficulty to the agent's current capabilities. Our system employs a controllable environment representation, extracts fine-grained performance feedback beyond binary success or failure, and implements a closed-loop adaptation mechanism that translates this feedback into environment modifications. This feedback-driven approach generates training environments that more challenging in the ways the agent needs to improve, enabling more efficient learning and better generalization to novel settings.

While DNN models have dramatically improved the accuracy of machine vision tasks, pervasive deployments of vision sensors for surveillance tasks continue to suffer from high energy consumption and network traffic overhead. To tackle these problems, we introduce CollabCam , an edge-based machine vision system designed for multi-camera deployments that leverages the naturally-occurring overlaps in the field-of-view (FoV) among neighboring cameras. CollabCam synergistically combines two innovative ideas: (a) having each individual camera compose and transmit mixed-resolution frames (MRF) via lightweight down-sampling, where the transmitted images have significantly lower resolution in the shared, overlapping portions , and (b) performing inference, for an exemplar object detection task, for each camera stream using a new collaborative mechanism which utilizes suitably-translated object bounding boxes from a peer “collaborating” camera as an additional input channel. We demonstrate how this collaborative mechanism is generalizable and can be realized by simply retraining off-the-shelf object detector DNNs, such as YOLOv3 and SSD, without modifying their model structures. By emulating the performance of CollabCam using two benchmark outdoor-campus multi-camera datasets, we show that Collab-DNNs can accommodate a 50–60 fold reduction in image size (therefore reducing network transmission overhead), for both high-resolution (1056x1056) and low-resolution (512x512) images, with a modest ≤ 2 - 5% drop in object detection accuracy, compared to a non-collaborative approach that suffers a ∼ 45–60% drop in accuracy. Subsequently, by deploying a Raspberry-Pi based CollabCam prototype on a campus-based test-bed, we demonstrate that CollabCam can reduce the overall energy/image frame overhead by ∼25–35%, with even higher energy savings (∼35–45%) likely with hardware optimization. Finally, additional experiments help demonstrate that CollabCam can prove beneficial for varied (including multi-class) object detection tasks and that CollabCam’s performance benefits may be best realized by ensuring that the number of deployed, collaborating cameras is not excessively high.

Future service robots have the potential to improve our quality of life across diverse contexts. However, they will introduce multifaceted privacy threats. After detailing such threats, we outline possible research approaches to elicit and implement user privacy preferences, considering robots’ capabilities and their deployment contexts, including multiuser scenarios in public spaces. Our work aims to catalyze future research efforts to promote the responsible, privacy-compliant deployment of such robots.

3-Dimensional Embodied Reference Understanding (3D-ERU) combines a language description and an accompanying pointing gesture to identify the most relevant target object in a 3D scene. Although prior work has explored pure language-based 3D grounding, there has been limited exploration of 3D-ERU, which also incorporates human pointing gestures. To address this gap, we introduce a data augmentation framework-Imputer, and use it to curate a new benchmark dataset-ImputeRefer for 3D-ERU, by incorporating human pointing gestures into existing 3D scene datasets that only contain language instructions. We also propose Ges3ViG, a novel model for 3D-ERU that achieves ~30% improvement in accuracy as compared to other 3D-ERU models and ~9% compared to other purely language-based 3D grounding models. Our code and dataset are available at https://github.com/AtharvMane/Ges3ViG.

We present a multimodal instruction comprehension framework, called MImIC, that utilizes visual sensing (including LIDAR and 2D RGB sensing) & AI spatial reasoning capabilities to support more seamless and immersive interaction between humans and AI-driven situated assistive agents. MImIC's key new capability is to support disambiguation of a wider set of relative spatial references that users naturally employ while issuing spatially-situated instructions. To support enhanced visual grounding via a combination of both fully-qualified and relative attribute references, MImIC uses (a) a fine-tuned transformer-based language translation DNN to accurately convert natural verbal commands into a structured set of machine understandable constraints (BLEU score=92.5), (b) a set of modules that use RGB+LIDAR sensing data to convert any relative attribute preferences to fully-qualified attribute constraints (median height/width estimation errors <=2cm), and (c) an enhanced image segmentation DNN, augmented with gesture+verbal cues, to extract target objects of interest (top-1 accuracy= ~85%). To demonstrate the viability and superiority of MImIC, we implement an exemplar AR-augmented, immersive furniture shopping application, called AIRFurn. AIRFurn allows users to browse for, select and overlay furniture items of interest using natural multi-modal and relative cues. experimental studies, using 34 & 11 users over 8 different layouts in a lab setting and 11 users in 6 different real-world home setups, show that AIRFurn offers superior performance, with significantly (~3x) lower task completion times, much higher task (17%+) accuracy and greater user satisfaction (SUS score= 78.8) compared to baselines where users perform selection using only fully-qualified verbal commands or manipulation of AR interfaces.