DLCL: Deep Learning-enabled Cooperative Localization for UAV-UGV Team with Adaptive Perception Recovery

International Journal of Computer Vision (IJCV) [under review] 

This paper presents a cooperative localization framework designed for heterogeneous UAV-UGV teams operating in GNSS-challenged environments. DLCL integrates deep learning-based multi-robot detection, an adaptive perception recovery mechanism, and GNSS-vision fusion to improve localization robustness. We introduce a new dataset of 30k UAV–UGV images and a 7.5k-frame sequential dataset for temporal modeling. A recurrent LSTM-based perception recovery (PR) module is proposed to predict future confidence trends and dynamically adjust detection thresholds, reducing frame-level failures and stabilizing temporal perception. Experiments using YOLOv8/v9/v10/v11 (nano-large) demonstrate that PR improves detection accuracy up to 50% across diverse trajectories, particularly under low illumination and small-object conditions. Ablation studies show that our GNSS-prior cropping strategy reduces computational load from 4.04→1.85 GFLOPs and inference latency from 36.74→21.86 ms, supporting real-time embedded deployment. Ablation studies further revealed that increasing the number of cooperating agents monotonically reduces global RMSE and tightens the 3σ uncertainty bounds, emphasizing the advantage of multi-robot cooperation. 

AnyThermal: Towards Learning Universal Representations for Thermal Perception

Accepted in ICRA 2026 [paper] 

We present AnyThermal, a thermal backbone that captures robust task-agnostic thermal features suitable for a variety of tasks such as cross-modal place recognition, thermal segmentation, and monocular depth estimation using thermal images. Our key insight is to distill the feature representations from visual foundation models such as DINOv2 into a thermal encoder using thermal data from these multiple environments. To bridge the diversity gap of the existing RGB- Thermal datasets, we introduce the TartanRGBT platform, the first open-source data collection platform with synced RGB- Thermal image acquisition.

Custom UAV with Model Predictive Control for Autonomous Static and Dynamic Trajectory Tracking in Agricultural Fields

Published at Frontiers in Robotics and AI [paper] 

This study introduces a custom-built uncrewed aerial vehicle (UAV) designed for precision agriculture, emphasizing modularity, adaptability, and affordability. Unlike commercial UAVs that are restricted by proprietary systems, the proposed platform integrates a Cube Blue flight controller for low-level control with a Raspberry Pi 4 companion computer that runs a Model Predictive Control (MPC) algorithm for high-level trajectory optimization. The system also incorporates Kalman filtering to enable adaptive mission planning and real-time coordination with a moving uncrewed ground vehicle (UGV), offering greater flexibility in changing field conditions. It consistently achieved root mean square error values between 8 and 20 centimeters during autonomous operations, with slightly higher errors in more complex trajectories. The UAV successfully followed the UGV along nonlinear, curved paths, confirming its suitability for real-world agricultural applications.

Cooperative Localization of UAVs in Multi-Robot Systems Using Deep Learning-Based Detection

Published at AIAA SCITECH 2025 Forum [paper] 

This paper introduces a novel cooperative localization framework designed to enhance localization accuracy in multi-robot systems comprising UAVs and Unmanned Ground Vehicles (UGVs). The proposed method leverages deep learning-based detection, specifically utilizing the YOLOv8 convolutional neural network, to enable real-time object detection and localization. By integrating perception with Kalman Filtering (KF), the approach achieves improved localization accuracy, even in challenging environments, thus advancing the state-of-the-art in cooperative multi-robot systems.

Digital soil mapping of available phosphorus using a smartphone-integrated RGB imaging device and ascorbic acid extraction method

Published at Smart Agricultural Technology [paper] 

To take quick remedial actions, it is critical to map the non-germinated mulch cells at a high throughput capacity which is aimed in this study. We compared the performance of different object detection models to find the missing index of mulch planting. 

Estimated the maximum resolution of testing images and calculated the maximum height for the drone flight. Targetting Computers and Electronics in Agriculture Journal.

A Two-stage Deep-learning Model for Detection and Occlusion-based Classification of Kashmiri Orchard Apples for Robotic Harvesting.

Published at Journal of Biosystems Engineering  [paper] 

Proposed a novel two-stage deep-learning-based approach that can detect the apples using YOLOv7 and utilized EfficientNet that can classify the apple's occlusion condition.

From Goals, Waypoints Paths To Long Term Human Trajectory Forecasting

Published at ReScience C Journal [paper] [code]

Undertaken as part of the Machine Learning Reproducibility Challenge 2021 , we reviewed the above-accepted ICCV 2021 publication for its reproducibility and verification of its empirical claims. Our major contributions include implementing new sampling methods, creating new visualizations, achieving better results than the original paper, and discovering the generalizing power of the model over other datasets.

Virtual Farm Environments and Sim-to-Real Transfer in Agricultural Robotics Using NVIDIA Omniverse

Selected at American Society of Agricultural and Biological Engineers Annual International Meet 2025

• Developed virtual farm environments in the NVIDIA Omniverse Isaac Sim to replicate real-world conditions like cornfields.

• Trained models using visual, force, and 3D data in simulation, the researchers aim to predict stalk lodging risk.

Robotic Manipulation for Plant Interaction Using Reinforcement Learning

Selected at Workshop on Machine Learning for Cyber-Agricultural Systema 2022 [GitHub] 

An agent was trained for bending the corn plant built in the "Chrono” simulator which calculates the finite element analysis. I did this project at Iowa State University.

Fine-tuning based approach for generalizing YOLOv4 network for Soybean detection in UAS images

Selected at American Society of Agricultural and Biological Engineers Annual International Meet 2022

• Explored two methods for achieving generalization of the YOLOv4 model using images collected from different fields with various real-life confusing differences like the presence of weeds in the field.

• Implemented Active Learning to reduce the burden of manually annotating the imagery dataset by 75 %.