← Back
All research projects
Projects
Work spanning 3D scene understanding and multimodal models, video world models, medical image analysis, and socially aware robot navigation.
Ground3D-LMM: Fine-Grained 3D Point Grounding & Spatial Reasoning
Sep 2025 – Mar 2026First author·Prof. Fahad Shahbaz Khan, MBZUAI·Accepted, ECCV 2026- Unified 3D LMM that, for the first time, jointly produces point-level 3D segmentation masks and metric-consistent numerical responses (size, distance, clearance) at object and part granularity.
- Proposed the 3D Grounded Measurement task and constructed Ground3D, a corpus of roughly 3M QA pairs across 2.5K ScanNet and ScanNet++ scenes with dense object and part annotations and multi-turn grounded dialogue.
- Achieves SOTA across object- and part-level grounding; +5.15% mIoU over Reason3D and ~20% improvement on ScanRefer instance grounding, with markedly lower metric error than image-only baselines.
- Released the Ground3D dataset publicly on Hugging Face, with over 13,000 downloads to date.
3D VisionLMMsGroundingSpatial ReasoningOpen-World Evaluation and Repair of Video World Models
Jun 2026 – PresentVisiting Researcher·Prof. Vineeth N. Balasubramanian, Microsoft Research India·Ongoing- Studying open-world failure modes in video world models, in particular object collapse for rare or unseen concepts under image-conditioned generation.
- Developing a detect–repair–re-evaluate pipeline that identifies collapsed generations, augments object references, and reconditions the model without retraining.
- Curating an object-centric benchmark spanning common, rare, and novel concepts to measure robustness across video world models.
Video GenerationWorld ModelsRobustnessBenchmarkingMicrocirculation Analysis with AI for Nailfold Capillaries
Aug 2024 – Dec 2024Lead developer·Prof. Tauhidur Rahman, MOSAIC Lab, UC San Diego·2 papers (KI Reports, Scientific Reports)- Built a deep-learning pipeline for nailfold capillary video analysis covering automated detection, tracking, and pattern classification linked to chronic kidney disease.
- 92% F1 using YOLOv11 + optical flow + graph-based methods, robust to artifacts and noisy clinical inputs.
- Shipped a Streamlit interface for real-time analysis, validated with collaborators at UCSD Health and Maastricht University.
Medical ImagingDetectionTrackingSANGO — Socially Aware Navigation through Grouped Obstacles
Jan 2024 – Aug 2024Second author·Prof. Sandeep Manjanna & Prof. Sunita Chauhan, Plaksha University·Published, ICC-10 2024- Deep RL method for socially compliant robot navigation: 83.5% reduction in social discomfort and 29.4% fewer collisions.
- Built MOSANG (Modular Socially Aware Navigation Gym), a custom simulation environment for training RL agents on real-world maps, integrating A* path planning and PPO via Stable Baselines with real-time visualization.
- 1st place in the SP Dutt Award for Innovation and Impact 2024 ($2,500), recognized for hospital-robot deployment potential.
Reinforcement LearningRoboticsSimulationSimMobility for Autonomous Rover Training
Jul 2023 – Aug 2023Pipeline developer·Prof. Malika Meghjani, MARVL Lab, SUTD·Research internship- Generalized SimMobility to simulate any OpenStreetMap-supported city worldwide for autonomous-vehicle training.
- End-to-end pipeline: OSM XML → PostgreSQL backend → modified C++ source for new road datasets.
Autonomous DrivingSimulationYotta Expo — Digital Marketplace for Street Vendors
Jul 2022 – Jun 2024Team lead·Prof. Rucha Joshi, Plaksha University·Deployed- Led a team of 3 to build a digital selling platform helping Indian street vendors raise revenue, cut waste, and compete with e-commerce.
- Shipped a three-app ecosystem (vendor, customer, delivery agent) to Google Play and iOS using Flutter and Firebase.
- Ran pilots in Prayagraj and Chandigarh: 20% increase in vendor revenue and 50% reduction in fruit and vegetable waste.
- Secured $1,000 annual funding and recognition from the US National Academy of Engineering.
MobileFlutterSocial Impact