News

Experience

 
 
 
 
 

Visiting Researcher

Carnegie Mellon University

Jun 2025 – Aug 2026 Pittsburgh, Pennsylvania, USA

  • Pursued my undergraduate thesis onsite at Carnegie Mellon University’s Robotics Institute under the supervision of Dr. Wenshan Wang and Dr. Ji Zhang from Jun 2025–Dec 2025, continued remotely since Jan 2026
  • Worked on 3D Object Representation Learning, Vision-Language Navigation for Outdoor Scenarios and co-organized the CMU-VLA Challenge

 
 
 
 
 

Research Intern

National University of Singapore

May 2024 – Dec 2025 Remote

  • Research intern at MARMoT Lab, NUS under the supervision of Prof. Guillaume Sartoretti
  • Worked on 3D Scene Graph Generation using Vision-Language Models
  • Designed a generalized framework for Open-Vocabulary 3D Gaussian Splatting, eliminating the need for a per-scene autoencoder
  • Developed a Next-Best-View Policy for 3D Reconstruction using Neural Radiance Fields (co-supervised by Prof. Marija Popović)

Publications

UAV_VLN

UAV-VLN: End-to-End Vision Language guided Navigation for UAVs
Pranav Saxena, Nishant Raghuvanshi, Neena Goveas
ECMR 2025  (Oral)
arXiv

UAV-VLN leverages the common-sense reasoning of LLMs and a vision model for cross-modal grounding to plan context-aware aerial trajectories from free-form natural language instructions.

LLM-RG

LLM-RG: Referential Grounding in Outdoor Scenarios using Large Language Models
Pranav Saxena, Avigyan Bhattacharya, Ji Zhang, Wenshan Wang
Human-aware Embodied AI Workshop @ IROS 2025  
arXiv, Code

A zero-shot hybrid pipeline that leverages Large Language Models for symbolic reasoning and Vision-Language Models for fine-grained attribute extraction to perform robust referential grounding in challenging outdoor driving scenarios.

3DSG

ZING-3D: Zero-shot Incremental 3D Scene Graphs via Vision-Language Models
Pranav Saxena, Jimmy Chiun
Multimodal Embodied Interaction in Robots Workshop @ ICRA 2026  
arXiv

Zero-shot framework for incremental 3D scene graph generation using vision-language models, enabling open-vocabulary reasoning and geometric grounding for embodied environments.

Gen-LangSplat

Gen-LangSplat: Generalized Language Gaussian Splatting with Pre-Trained Feature Compression
Pranav Saxena
Preprint  
arXiv, Code

A generalized 3D Gaussian Splatting framework that eliminates per-scene training and improves cross-scene generalization via a ScanNet-trained autoencoder for CLIP feature compression. This work was done in collaboration with NUS and TU Delft.

Other Projects

State Tracking

Persistent Object Identity and Tracking Through State Transformations
University of Edinburgh

Developing an online, multi-object tracking framework for egocentric videos that follows objects through state transformations.

3D Representation

3D Object Representation Learning
Carnegie Mellon University

Learning generalized 3D object representations from point clouds and multi-view visual observations for downstream 3D understanding and reconstruction tasks.

NBV-NeRF

Next-Best-View Policy for NeRFs
National University of Singapore, TU Delft, Purdue University
Code

Developed an attention-based Next-Best-View (NBV) policy for Neural Radiance Fields to select the most informative viewpoints for 3D reconstruction. Used a Vision Transformer to encode multi-view image embeddings and candidate poses to predict ΔPSNR gain, supervised with ground-truth from pixelNeRF.

wifi-Perception

3D Human Pose Estimation using WiFi signals
BITS Pilani, Goa
Report

Developed a two-stage learning framework for 3D human pose estimation using WiFi signals, and incorporated phase sanitization to improve CSI quality. Validated on the MM-Fi dataset, achieving enhanced accuracy and robustness over raw CSI inputs.

Artemis

Artemis
BITS Pilani, Goa
Video

Developed an autonomous drone using ROS2 and MAVROS as part of my competitive autonomous drone team, Team Artemis. Implemented a robust surveying and informative path planning algorithm. Hardware stack includes a Raspberry Pi 5, PX4 flight controller.

Swarm Bots

Swarm Bots
BITS Pilani, Goa
Poster

Developed an algorithm for swarm robots by combining Artificial Potential Field and Reciprocal Velocity Obstacles methods to overcome the shortcomings of both methods.

Contact