- We test how images, instructions, and visual memory causally affect four VLM-based navigation policies.
- We then test whether navigation concepts and explicit reasoning remain steerable after fine-tuning and visual shift.

- Designed a network to predict the scaling factor for generative navigation policies to create metric waypoints in the real-world
- Used the real-world waypoints to guide the robot away from obstacles and towards the goal beating SOTA.

- Used Conditional Flow Matching(CFM) along with depth embeddings for efficient goal-conditioned robot navigation as opposed to diffusion policies
- Achieved the higher accuracy and runtime efficiency compared to the SOTA.

- Solved data association for underwater SLAM through a novel method for sonar image correspondence using Learned Features.
- Introduced a pose-supervised network that generates feature descriptors robust to changes in viewpoints, enabling more reliable feature matches in sonar based localization and mapping.

- Developed a novel audio generation framework (AADG) using LLMs and text-to-audio models to simulate realistic scenarios specifically for Audio Anomaly Detection and Localization.
- Expanded anomaly detection datasets beyond industrial sounds, addressing diverse real-world environments like telephonic or video-derived audio.
- Created a modular, rigorously verified framework enabling scalable and reliable generation of realistic audio data, improving robustness of anomaly detection models against out-of-distribution scenarios.

- Finetuned Longformer on a scraped Wikipedia dataset to find most important tokens based on attention score.
- Used other conventional techniques such as Word2Vec, Doc2Vec and BERT based models understand words which lead to polarisation.
- Paper accepcted at AAAI 2023 Workshop on AI for credible elections and selected for publication at the AI Magazine Fall 2023.

- Implemented on Object based semantic SLAM, created low-memory metric semantic maps for multi-robot communication.
- Formulated a graph based SLAM. Used quadric factors with the factor graph with underlying visual inertial odometry.
- Designed feature descriptors for SONAR using unsupervised learning for underwater SLAM.
- Used an encoder decoder structure with CNNs with custom loss functions to learn without labels.

- Developed an algorithm for detecting and localizing multiple leaks in Water Distribution Systems with Intermittent water supply for Bengaluru a city of 8.5mil.(As part of EqWATER funded by Ministry of Human Resources Development, Govern- ment of India).
- Automated intermittent water supply in an experimental system(60m long*100mm dia network, 4L/s) using LabVIEW to generate scaled-down comparable data with identical disturbances and leaks analogous to field data.