Pomelo: Multi-Perspective Generative Model
Public research demo · 2026
A flexible generative model that can solve multiple tasks in one go: novel view synthesis, 3DGS reconstruction, camera-controlled flythroughs and 360s.
Nikhil Keetha
Howdy! I'm a final-year Ph.D. candidate at Carnegie Mellon University, advised by Sebastian Scherer and Deva Ramanan. I'm also a Visiting Researcher at Meta, advised by Peter Kontschieder.
I work on generalizable spatial perception aka learning representations that let embodied systems perceive, understand, and reason about the physical world. My research spans large-scale learning, visual correspondence, 3D/4D understanding, and spatial memory, with applications in field robotics, augmented reality, and interactive agents.
I received my M.S. in Robotics from CMU and B.Tech. in Engineering Physics from IIT (ISM) Dhanbad. Previously, I interned at Mila and the QUT Centre for Robotics, and was a CMU Robotics Institute Summer Scholar. I've had the pleasure of collaborating with many brilliant friends, collaborators and mentors!
Outside research, I enjoy movies, cooking, music, anime, F1, sports, gaming and racing.
Featured Research
Selected work, newest first. Full publication list on Google Scholar.
Any4D: Unified Feed-Forward Metric 4D Reconstruction
CVPR 2026
A feed-forward model that jointly predicts metric 3D geometry and dense motion across multiple frames, with flexible support for additional sensor inputs.
RAVEN: Resilient Aerial Navigation via Open-Set Semantic Memory and Behavior Adaptation
ICRA 2026 · Oral
Persistent 3D semantic memory and adaptive search behaviors enable aerial robots to find objects across large, unstructured outdoor environments.
Demonstrating ViSafe: Vision-enabled Safety for High-speed Detect and Avoid
RSS 2025 · Demonstrations
A vision-only detect-and-avoid system which integrates high-speed perception with safety-aware control. Demonstrated real flight tests with closure rates up to 144 km/h.
RayFronts: Open-Set Semantic Ray Frontiers for Online Scene Understanding and Exploration
IROS 2025
An open-set semantic mapping system which combines 3D voxels with ray frontiers, letting robots reason about and explore beyond their depth-sensing range.
FlowR: Flowing from Sparse to Dense 3D Reconstructions
ICCV 2025 · Highlight
Learning a direct mapping between incorrect renderings and their corresponding ground-truth images, augmenting scene captures with consistent novel, generated views to improve reconstruction quality.
Meta Hyperscape
Public demo · 2024 / Capture launch · 2025
Photorealistic reconstruction turning captured physical spaces into immersive digital replicas that people can explore and share in VR.
Map It Anywhere (MIA): Empowering Bird's Eye View Mapping using Large-scale Public Data
NeurIPS 2024
A data engine pairing street-level imagery with public maps at scale, enabling more generalizable bird’s-eye-view mapping across diverse environments.
SplaTAM: Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM
CVPR 2024
Explicit 3D Gaussians enable online tracking and dense mapping from a single unposed RGB-D camera, with high-quality scene reconstruction and rendering.
FoundLoc: Vision-based Onboard Aerial Localization in the Wild
arXiv 2023
AnyLoc-based place recognition and visual-inertial odometry enable onboard aerial localization against satellite imagery, without GPS or a known starting pose.
AnyLoc: Towards Universal Visual Place Recognition
IEEE RA-L 2024 · ICRA 2024
Off-the-shelf image foundation model features and unsupervised aggregation enable state-of-the-art visual place recognition across environments, without task-specific training.
A few works are highlighted. * indicates equal contribution.
Awards
- 2025–2027
Meta Academic Research Grants
Thesis research at CMU supported by two Meta Academic Research Grants ($400K each).
- 2026
ECCV 2026
- 2026
CVPR 2026
- 2025
ICCV 2025
- 2025
CVPR 2025
- 2021
Robotics Institute Summer Scholar
Carnegie Mellon University
- 2020
ACCV 2020
Press
Meta Hyperscape
ViSafe
Map It Anywhere
Talks
- Jun 2026
Image Matching: Local Features & Beyond
Invited workshop talk · CVPR 2026
- Nov 2025
- Fall 2025
Guest lecture · CMU 16-820
- Fall 2025
- Fall 2024
- Spring 2024
Service
Reviewing
- Conferences and workshops: CVPR, ICCV, ECCV, NeurIPS, ICRA, IROS, 3DV, and the Image Matching Workshop (IMW).
- Journal: IEEE Robotics and Automation Letters (RA-L).
Teaching
- Teaching Assistant, 16-824: Visual Learning and Recognition, CMU · Spring 2025.
- Teaching Assistant, 16-820: Advanced Computer Vision, CMU · Fall 2024.
Admissions
- CMU Robotics Institute Graduate Admissions Committee.
- CMU Robotics Institute Summer Scholars (RISS) Admissions Committee.
Mentorship
- Mentored many researchers at CMU and research interns at Meta.
- CMU Undergraduate AI Mentoring Program and SCS Mentorship Program.
Outreach
RoboLaunch
Co-organized a robotics outreach initiative broadening participation and making robotics more accessible. The 2022 series reached over 13,000 students and viewers worldwide.
Tartan Planning Series
Co-organized the Spring 2023 series of talks, tutorials, and interactive learning on robot planning.
Tartan SLAM Series
Co-organized an interactive series of talks, tutorials, and learning on simultaneous localization and mapping.
Tartan SLAM Series — Fall Edition
Co-organized the Fall 2021 edition, connecting the robotics community through talks and tutorials on SLAM.
Projects
Research collaborations
At CMU, I have worked on collaborative research with these partners and programs:
- AIST — Frontier AI Research, Japan
- IARPA — WRIVA
- Near Earth Autonomy
- DARPA — Shared Airspace
- DSTA — Multi-Robot Search & Rescue
- Bosch
Earlier software projects
- GradSLAM RGB-D Completion
Leveraged Multi-view gradients from GradSLAM to optimize RGB-D Images. - COVID-19 Twitter Sentiment Analysis India IBM Hackathon 2020
Developed a Visualization Dashboard for COVID-19 Twitter Sentiment Analysis & Extraction using a modified Roberta & Roberta-CNN model - Twitter Sentiment Extraction Kaggle Competition
Developed a modified Roberta base model with CNN head for Twitter Sentiment Extraction & Implemented Teacher Student learning and Ensembling for improvement of Model Performance. Placed 58/2,227 (Top 3%) - Silver Medal - Reddit Flair Predictor
Developed an XLNet language model based on Transfer Learning & a Web Application based on Flask API for Reddit Flair Classification