Research

The central theme of this research thread is the non-invasive detection of disease biomarkers, particularly volatile organic compounds (VOCs) exhaled in human breath, as a painless alternative to finger prick-based diagnostics. Key efforts include:

• Breath-based blood glucose monitoring — sensor arrays combined with machine learning to predict blood glucose levels from exhaled acetone and other VOC markers.
• Clinical data collection platforms — custom embedded platform for breath VOC data collection in the wild. The platform is IRB human-subject study compatible and has been tested and verified. The patented device and companion smart phone app transmits data to the researchers.
• On-device machine learning — lightweight ML models for real-time inference.

Sim-to-real transfer learning, reinforcement learning for robotic arms, imitation learning, and vision-guided control. Our robots learn from simulation and human demonstration, then operate on physical systems.

• Imitation learning for bimanual coordination — using Universal Manipulation Interface (UMI) to collect human demonstration data and train policies for two-arm handoff tasks.
• Sim-to-real reinforcement learning — training robot arm control policies in simulation (GR00T, PyBullet, IsaacGym) and transferring to real hardware with minimal real-world fine-tuning.
• Vision-Language-Action (VLA) systems — integrating large language models and visual perception for task-conditioned robot control.
• NLP-guided navigation — natural language command parsing for autonomous mobile robot navigation in indoor environments.

A horizontal thread across all ISL projects: building the embedded hardware and firmware that deploys intelligent algorithms outside the lab, on constrained devices:

• Sensor acquisition systems — custom PCB designs for gas sensors, biosensors, and multi-modal arrays with analog front-end conditioning and ADC integration.
• RTOS-based firmware — FreeRTOS implementations on STM32, ESP32, and ARM Cortex-M platforms for real-time data acquisition and BLE/WiFi telemetry.
• TinyML / Edge inference — TensorFlow Lite and ONNX model deployment on microcontrollers for on-device classification and regression.
• Environmental monitoring — VOC, particulate matter, and CO₂ sensing networks for indoor air quality research.

This research thread bridges visual perception and language understanding to build autonomous systems capable of interpreting the world and acting on instructions:

• Medical image classification — deep learning models for diabetic retinopathy detection from fundus images, targeting resource-constrained clinical settings.
• Vision-Language-Action (VLA) integration — combining visual encoders and large language models to enable task-conditioned robot manipulation from natural language commands.
• Autonomous navigation via NLP — natural language command parsing for indoor mobile robot navigation, mapping instructions to motion planning in real environments.
• Object detection and tracking — real-time vision pipelines for robot arm manipulation, grasp planning, and workspace monitoring.