Charuvahan Adhivarahan
Multimodal Perception for Embodied Intelligence
Postdoctoral Associate
University at Buffalo
Table of Contents
Research Interests
Robots that operate beyond controlled laboratories need representations that remain useful as their sensors, surroundings, and embodiments change. My research builds multimodal representations and world models that integrate vision, depth, and wireless signals into compact internal models that robots can update and reason over in real time. I study how these representations can preserve relevant structure when sensor configurations change, environments shift, or knowledge must transfer between robot embodiments. The goal is perception that supports sustained autonomy in the varied and unpredictable conditions of real deployments.
Projects
Gaussian Splatting the Physical World
Robots need world representations that preserve appearance and geometry without becoming too expensive to update, share, or use for action. Gaussian splats are a promising foundation: they combine photorealistic rendering, geometric fidelity, and greater sample efficiency than point clouds. Turning them into a robust perceptual backbone, however, requires solving several open problems. Their smooth primitives struggle to preserve the sharp boundaries needed for contact-rich manipulation; incremental online updates remain underdeveloped; distributed multi-robot pipelines require reliable correspondence and consistency; and current computational and memory costs remain prohibitive for real-time deployment.
Looking Beyond the Visible Spectrum
Reliable plastic recycling depends on measurements that can distinguish materials conventional cameras cannot. Our research combines infrared light, thermal signatures, electrical charge, and dielectric fields with machine learning to recover information hidden within plastic materials. These non-destructive measurements identify plastic types with up to 100% accuracy and quantify recycled content with over 97% precision. By grounding perception in molecular fingerprints and thermal behavior, this work turns otherwise invisible physical signals into information that recycling systems can use.
Sensing the World at Every Scale
No single sensor or vantage point captures enough of a complex environment to support reliable understanding. Our research integrates visible, multispectral, near-infrared, shortwave-infrared, and thermal observations collected from orbital satellites, UAV surveys, and ground-level UGV scans. Mutual-information-based learning allows models to exchange structure across modalities and resolutions rather than treating each data source in isolation. The resulting representations combine complementary evidence across wavelength and scale, building richer models for robots and sensing systems that must reason about large, heterogeneous environments.
Publications
- Yaoli Zhao, Charuvahan Adhivarahan, Chandra Lekha Jyothula, Karthik Dantu, Thomas Thundat, and Amit Goyal, Determining the Percentage of Recycled Plastic Content in a Plastic Product, Communications Engineering, Vol. 5, No. 51, 2026.
- Charuvahan Adhivarahan and Karthik Dantu, WISDOM: WIreless Sensing-assisted Distributed Online Mapping, 2019 International Conference on Robotics and Automation (ICRA), 2019, pp. 8026–8033
- Zakieh S. Hashemifar, Charuvahan Adhivarahan, Anand Balakrishnan, and Karthik Dantu, Augmenting Visual SLAM with Wi-Fi Sensing for Indoor Applications, Autonomous Robots, Vol. 43, No. 8, 2019, pp. 2245–2260
- Charuvahan Adhivarahan, Zakieh Sadat Hashemifar, and Karthik Dantu, Improving RGB-D SLAM using Wi-Fi, 16th International Conference on Information Processing in Sensor Networks (IPSN), 2017
- Long Duong, Charuvahan Adhivarahan, Roshan Ayyalasomayajula, and Karthik Dantu. 2026. TIPS: Thermal Image based Plastics Sorting. In The 24th ACM International Conference on Mobile Systems, Applications, and Services (MobiSys ’26), June 21–25, 2026, Cambridge, UK. ACM, New York, NY, USA. (Accepted)
- Zaid Tasneem, Charuvahan Adhivarahan, Dingkang Wang, Huikai Xie, Karthik Dantu, and Sanjeev J. Koppal, Adaptive Fovea for Scanning Depth Sensors, The International Journal of Robotics Research, Vol. 39, No. 7, 2020, pp. 837–855
- Stephen Xia, Minghui Zhao, Charuvahan Adhivarahan, Kaiyuan Hou, Yuyang Chen, Jingping Nie, Eugene Wu, Karthik Dantu, and Xiaofan Jiang, Anemoi: A Low-cost Sensorless Indoor Drone System for Automatic Mapping of 3D Airflow Fields, Proceedings of the 29th Annual International Conference on Mobile Computing and Networking (ACM MobiCom '23), Article No. 77, 2023
- Yash Turkar, Pranay Meshram, Christo Aluckal, Charuvahan Adhivarahan, and Karthik Dantu, Empir3D: A Framework for Multi-Dimensional Point Cloud Assessment, arXiv preprint arXiv:2306.03660, 2023
- Yash Turkar, Christo Aluckal, Charuvahan Adhivarahan, Alessandro Sebastiani, and Karthik Dantu, A View-Planning Approach to 3D Reconstruction, IEEE International Conference on Extended Reality (XR), 2024, pp. 340–350
- Yash Turkar, Shaunak De, Charuvahan Adhivarahan, Luca Mottola, Alessandro Sebastiani, Davide Castelletti, and Karthik Dantu, Enhancing Archaeological Surveys with InSAR Imagery and UAV-Based GPR, IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 2024, pp. 11454–11456
- Kartikeya Singh, Yash Turkar, Christo Aluckal, Charuvahan Adhivarahan, and Karthik Dantu, PANOS: Payload-Aware Navigation in Offroad Scenarios, arXiv preprint arXiv:2409.16566, 2024
- Shiv Mehta, Vaishali Maheshkar, Charuvahan Adhivarahan, and Karthik Dantu, CLIPS: Continual Learning Infrastructure for Plastics Sorting, 2025 International Conference on Machine Learning and Applications (ICMLA), 2025, pp. 1197–1204
- Pranay Meshram, Yash Turkar, Kartikeya Singh, and Praveen Raj Masilamani, QAL: A Loss for Recall-Precision Balance in 3D Reconstruction, Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2026
- Vaishali Maheshkar, Aadarsh Anantha Ramakrishnan, Charuvahan Adhivarahan, and Karthik Dantu, Detailed Evaluation of Modern Machine Learning Approaches for Optic Plastics Sorting, arXiv preprint arXiv:2505.16513, 2025
- Pranay Meshram, Charuvahan Adhivarahan, Ehsan Tarkesh Esfahani, Souma Chowdhury, Chen Wang, and Karthik Dantu, CLEAR: A Semantic-Geometric Terrain Abstraction for Large-Scale Unstructured Environments, arXiv preprint arXiv:2601.13361, 2026
Education
University at Buffalo, Buffalo, New York, USA
Ph.D., Computer Science, May 2023
Advisor: Dr. Karthik Dantu
M.S., Computer Science, May 2018
Annamalai University, Chidambaram, Tamilnadu, India
B.E., Computer Science, May, 2009
Academic Experience
University at Buffalo, Buffalo, New York, USA
Postdoctoral Associate, Computer Science. (Current)
Novel representations for Embodied Intelligence with multi-modal sensing.
Research Assistant January, 2018 - Aug 2023
Duties at various times have included research into high fidelity motion capture and maintaining the SMART Motion Capture Lab, training users on the mocap system and robots like the Baxter, URX arms and the Husky and consulting in design of experiment for motion capture.
Teaching Assistant August, 2017 - December, 2017
Duties at various times have included conducting office hours and recitations for CSE 468/568 Intro to Robotics Algorithms and CSE 487/587 Data Intensive computing. Topics include: kinematics, probabilistic algorithms for localization and mapping, planning, and navigation for Robotics Algorithms and MapReduce and predictive analytics, statistical software packages in R and Python and big data infrastructures like Hadoop and Spark ecosystems for Data Intensive Computing
Professional Experience
HCL Technologies, Chennai, Tamilnadu India
Senior Software Engineer Jan, 2010 - Feb, 2014
Engineered several projects, including end-to-end development and maintainance of shopping application with database design, web services, front-ends for the web, phone and television
Contact Information
Computer Skills
- Frameworks: ROS, Tensorflow, PyTorch
- Python, C, C++, C\#, Java, JavaScript
- Algorithms: Simultaneous Localization and Mapping packages, Task Allocation, Multi-agent Task Planning and Reinforcement Learning
- Operating Systems: GNU/Linux, Windows and MacOS