Pranay Meshram

FIELD ROBOTICS · PLANNING · 3D PERCEPTION

I build representations and planners that help robots reason across large, unstructured environments—from terrain semantics and geometry to mission language and robot-specific constraints.

Explore the work Publications Résumé

I am a Ph.D. student at the University at Buffalo, advised by Dr. Karthik Dantu in the DRONES Lab. My research connects language-conditioned planning, semantic-geometric terrain abstraction, hierarchical search, and 3D perception so autonomy systems can make fast, grounded decisions outside structured roads.

My current work develops robot-adaptive off-road planning from natural-language mission preferences. Earlier work spans HOPHY and CLEAR for large-scale terrain reasoning, QAL and Empir3D for point-cloud reconstruction quality, modular visual SLAM, and self-supervised stereo depth from polarization at Meta Reality Labs.

SELECTED WORK

From terrain maps to field robots

View the full portfolio

Two-stage language-conditioned route-planning framework

2026 · Ongoing research

Language-Conditioned Off-Road Planning

Robot-adaptive route planning that grounds mission preferences in semantic terrain and traversal consequences.

Two-stage region-guided planning reduces search states by 75% and candidate latency by 20–22% in the current evaluation.

2026 · arXiv 2609.20694 · Under submission

HOPHY

Hierarchical off-road planning with reusable hypergraph abstractions for fast, localized replanning.

68–1,280× faster path queries; validated over 1.5 km of autonomous Jackal navigation.

2026 · IEEE RA-L

CLEAR

A semantic-geometric terrain abstraction for reliable planning across large, unstructured environments.

10× faster planning with 6.7% cost overhead in the reported evaluation.

QAL coverage and attraction terms for point-cloud reconstruction

2026 · WACV

QAL

A quality-aware loss that separates coverage and precision for point-cloud reconstruction.

+4.3 average coverage points over Chamfer Distance in the reported experiments.

2026 · CRV

Empir3D

Multi-dimensional point-cloud assessment across resolution, accuracy, coverage, and artifacts.

A four-dimensional view of reconstruction quality beyond a single distance score.

news

Aug 06, 2026	Our paper “CLEAR: A Semantic-Geometric Terrain Abstraction for Large-Scale Unstructured Environments” has been accepted to IEEE Robotics and Automation Letters (RA-L). For project details, visit the CLEAR project page.
Dec 09, 2025	Our paper “QAL: A Loss for Recall–Precision Balance in 3D Reconstruction” has been accepted to WACV 2026! 🎉
Nov 02, 2025	UBPercept placed 5th overall at the 2022 ACM/IEEE TinyML Design Contest at ICCAD (Nov. 2), and earned 1st in latency and 3rd in flash (memory). Full results: TinyML Contest Winners.
Jul 08, 2022	Built a monocular visual odometry (VO) pipeline in Rust on the KITTI benchmark. Implemented feature tracking, pose estimation, and trajectory reconstruction, with plotting utilities to compare estimated trajectories against ground truth. Benchmarked VO accuracy across sequences and experimented with tuning feature detection thresholds and RANSAC parameters to handle motion blur and texture-poor regions. Added demos and visualizations to highlight drift over long trajectories and loop-closure opportunities. Git Repository
Jun 21, 2022	DAC System Design Contest 2022 – 4th place (UBPercept). Built an FPGA-friendly CNN pipeline on Ultra96V2 with quantization-aware training and deployment automation, targeting tight latency/memory budgets. Led a 7-member team through data curation, profiling, and inference optimization; automated builds and on-board evaluation to iterate rapidly on accuracy–efficiency trade-offs. Documented lessons on model pruning and kernel fusion for edge devices. Results

selected publications

arXiv
HOPHY: A Hierarchical Hypergraph Representation for Off-Road Path and Mission Planning

Pranay Meshram, Charuvahan Adhivarahan, Prithvi Poddar, and 4 more authors

arXiv preprint arXiv:2609.20694, Sep 2026

Under submission for IEEE Journal

Abs arXiv Bib HTML

HOPHY introduces a reusable hierarchical terrain representation for repeated off-road path and mission planning. It organizes terrain into geometrically connected semantic regions, connectivity-preserving coarse regions, and typed hyperedges for changing terrain, agent, and weather context. The evaluation reports reliable planning with near-oracle path costs, lower query and replanning latency, and a 1.5-km physical Clearpath Jackal mission.
@article{meshram2026hophy, title = {HOPHY: A Hierarchical Hypergraph Representation for Off-Road Path and Mission Planning}, author = {Meshram, Pranay and Adhivarahan, Charuvahan and Poddar, Prithvi and Esfahani, Ehsan Tarkesh and Wang, Chen and Chowdhury, Souma and Dantu, Karthik}, journal = {arXiv preprint arXiv:2609.20694}, year = {2026}, month = sep, note = {Under submission for IEEE Journal}, webpage = {https://droneslab.github.io/HOPHY/}, }
RA-L
CLEAR: A Semantic-Geometric Terrain Abstraction for Large-Scale Unstructured Environments

Pranay Meshram, Charuvahan Adhivarahan, Ehsan Tarkesh Esfahani, and 3 more authors

IEEE Robotics and Automation Letters, Sep 2026

Abs arXiv Bib HTML

Long-horizon navigation in unstructured environments demands terrain abstractions that scale to tens of km\textsuperscript2 while preserving semantic and geometric structure, which neither uniform grids nor quadtrees provide. CLEAR combines boundary-aware spatial decomposition with recursive plane fitting to produce convex, semantically aligned regions encoded as a terrain-aware graph, enabling faster and more reliable planning on large areas.
@article{meshram2026clear, title = {CLEAR: A Semantic-Geometric Terrain Abstraction for Large-Scale Unstructured Environments}, author = {Meshram, Pranay and Adhivarahan, Charuvahan and Esfahani, Ehsan Tarkesh and Chowdhury, Souma and Wang, Chen and Dantu, Karthik}, journal = {IEEE Robotics and Automation Letters}, year = {2026}, webpage = {https://droneslab.github.io/CLEAR/}, }
WACV
QAL: A Loss for Recall–Precision Balance in 3D Reconstruction

Pranay Meshram, Yash Turkar, Kartikeya Singh, and 3 more authors

In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Mar 2026

Abs DOI arXiv Bib HTML

Volumetric learning underpins many 3D vision tasks such as completion, reconstruction, and mesh generation, yet training objectives still rely on Chamfer Distance (CD) or Earth Mover’s Distance (EMD), which fail to balance recall and precision. We propose Quality-Aware Loss (QAL), a drop-in replacement for CD/EMD that combines a coverage-weighted nearest-neighbor term with an uncovered–ground-truth attraction term, explicitly decoupling recall and precision into tunable components.
@inproceedings{meshram2026qal, title = {QAL: A Loss for Recall--Precision Balance in 3D Reconstruction}, author = {Meshram, Pranay and Turkar, Yash and Singh, Kartikeya and Masilamani, Praveen Raj and Adhivarahan, Charuvahan and Dantu, Karthik}, booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)}, pages = {7965--7974}, month = mar, year = {2026}, doi = {10.1109/WACV61042.2026.00769}, }
CRV
Empir3D: A Framework for Multi-Dimensional Point Cloud Assessment

Yash Turkar, Pranay Meshram, Christo Aluckal, and 2 more authors

In Proceedings of the 23rd Conference on Robots and Vision (CRV), May 2026

Abs DOI arXiv Bib HTML

Empir3D proposes a multi-dimensional point cloud evaluation framework with metrics for resolution, accuracy, coverage, and artifacts, providing richer assessment of SLAM and point cloud completion results beyond traditional Chamfer or EMD.
@inproceedings{turkar2026empir3d, title = {Empir3D: A Framework for Multi-Dimensional Point Cloud Assessment}, author = {Turkar, Yash and Meshram, Pranay and Aluckal, Christo and Adhivarahan, Charuvahan and Dantu, Karthik}, booktitle = {Proceedings of the 23rd Conference on Robots and Vision (CRV)}, month = may, year = {2026}, doi = {10.21428/d82e957c.900da59a}, }