Projects
Research Projects
MapForest: A Modular Field Robotics System for Forest Mapping and Invasive Species Localization (Sep 2024 – Present) MS Thesis — Carnegie Mellon University, Kantor Lab
Monitoring invasive tree species across large forests is challenging due to limited accessibility, manual scouting reliance, and degraded under-canopy GNSS. MapForest is a modular field robotics system that transforms multi-modal sensor data (LiDAR, IMU, GNSS, RGB) into GIS-ready invasive-species maps. The system features a compact, platform-agnostic sensing payload deployable on UAV, bicycle, or backpack platforms. The software pipeline combines LiDAR–inertial SLAM (GLIM with covariance-aware GNSS factors and robust loss kernels) with a YOLOv8-based Tree-of-Heaven (Ailanthus altissima) detector, fusing detections into georeferenced GeoTIFF layers for downstream GIS analysis. Evaluated across six sites spanning urban, park, trail, and forest environments: trajectory deviation error of 1.95 m over a 1.2 km forest traversal, Tree-of-Heaven detection F1 of 0.77. Datasets and tooling are publicly released.
Authors: Sandeep S. Zachariah, Francisco Yandun, Sachet Korada, Abhisesh Silwal
[arXiv] [Website]
Dense Forest Under-Canopy and Above-Canopy Map Merging (May 2025 – Present) MS Thesis — Carnegie Mellon University, Kantor Lab
Foliage occlusions prevent complete 3D reconstruction of dense forest environments from any single viewpoint. This work develops a method to align under-canopy (terrestrial LiDAR/mobile) and above-canopy (UAV/aerial) maps by maximizing mutual information in a shared latent space of tree-likelihood fields inferred from each map. The approach targets dense broadleaf forests where viewpoint-induced appearance differences make correspondence-driven registration unreliable.
Fire Blight Detection in Apple Orchards (Jan 2025 – May 2025) Field Robotics Course Project — Carnegie Mellon University
Erwin is a fully autonomous robotic inspection system for detecting and mapping fire blight (Erwinia amylovora) in high-density apple orchards during dormant season. The system uses an Amiga mobile base with a 6-DOF xArm manipulator and custom stereo vision sensors. I led system integration and navigation, implementing the mission control software as a behaviour tree in ROS2 to orchestrate navigation (RTK-GNSS waypoint following, obstacle avoidance), disease detection (stereo-vision semantic point cloud pipeline), manipulation (collision-free arm trajectories for full-tree coverage), and spray marking of infected trees. Simulation was performed in NVIDIA Isaac Sim. The Foxglove-based UI provides a georeferenced orchard map with infected trees overlaid in real time. Field-tested at Penn State Fruit Research and Extension Center.
Team: Jack Nelson, Daniya Nussipbek, Sarthak Jain, Hayden Feddock, Sandeep S. Zachariah, Yi Wu
Leveraging Large Language Models for Robotic Systems (May 2023 – June 2024) Indian Institute of Technology Delhi
Current robotic systems are limited to interaction with a predefined set of objects. This project leveraged the zero-shot generalization capabilities of LLMs and VLMs — trained on internet-scale data — to create a holistic robotic system encompassing open-world perception, task planning, and low-level behaviors. Developed an open-world scene graph representation that admits local updates as the scene evolves, achieving superior performance on open-world object detection and relation extraction compared to prior methods.
[Paper] [Website]
Task Planning under Uncertainty for Language-Guided Mobile Manipulator (July 2022 – August 2023) Indian Institute of Technology Delhi
A robot working alongside a human must comprehend natural language instructions, perceive the environment, and generate action sequences in uncertain, partially observable outdoor settings. Our system monitors execution of low-level actions and replans on any discrepancy between expected and perceived state. For partial observability, the planner spawns contextually grounded exploratory behaviors, achieving a Goal Reaching Rate (GRR) of 90%.
[Video] [Manuscript]
Underwater Image Enhancement (July 2021 – April 2022) Undergraduate Thesis — NIT Calicut
Developed a novel wavelet-fusion method combining Multiscale Retinex and Dark Channel Prior (DCP) with custom color balancing algorithms for underwater image enhancement. The method outperformed both naive approaches and learning-based methods (U-Net, CycleGAN) on quantitative and qualitative metrics. Superiority was demonstrated through improved accuracy in underwater object detection.
[Thesis]
Modelling, Control, and Simulation of Quadruped Robots (April 2021 – September 2021) Project Intern — IIST
Kinematic and dynamic modelling of a quadruped robot using Composite Rigid Body and Recursive Newton-Euler algorithms. Implemented closed-loop inverse kinematics (CLIK) for 6D pose control of torso and foot frames. Joint-space and task-space inverse dynamics control, incorporating holonomic and non-holonomic constraints, contact impact dynamics, and constraint drift compensation, all simulated in MATLAB.
[Video 1] [Video 2] [Report]
Modelling, Control, and Simulation of a Planar Robotic Manipulator (May 2019 – July 2019) Summer Intern — IIST
Joint-space and task-space kinematic and dynamic analysis of a 2R manipulator via Euler-Lagrange equations. CAD model created in SolidWorks; mass-inertia properties fed into MATLAB/SIMULINK control design. Structural dynamic analysis in ANSYS to verify absence of control-structure interaction.
[Report]
Course Projects
WanPolicy: Representation Alignment for Robust and Efficient World-Action Models (Jan 2026 – Present) Carnegie Mellon University — NeurIPS 2026 Submission
World-Action Models (WAMs) repurpose pretrained video diffusion backbones for robot manipulation by jointly modeling future observations and actions. WanPolicy challenges the dominant pixel-level reconstruction paradigm, replacing the diffusion objective with a representation alignment loss that supervises intermediate DiT features using a frozen DINOv3 encoder. Fine-tuning the DiT as a regressor eliminates iterative denoising at inference — a single forward pass suffices, yielding significant speedups. On LIBERO-Plus (environmental perturbations), WanPolicy achieves 80% success vs. the prior art of 69%, while training on substantially less robot data than competing methods.
Semantic Simultaneous Localization and Mapping (Jan 2025 – Apr 2025) Carnegie Mellon University
Collected a semantic SLAM benchmark on the CMU campus using a custom sensor suite (LiDAR, camera, IMU, GNSS). Evaluated state-of-the-art semantic SLAM systems — Kimera, SegMap, and SlideSLAM — on the benchmark, analysing metric and semantic accuracy across diverse campus environments.
Diffusion Policy and Flow Matching for Long-Horizon Manipulation (Aug 2025 – Dec 2025) 16-831: Introduction to Robot Learning — Carnegie Mellon University
Unified evaluation of diffusion-based and flow-matching action-sequence policies on the RoboMimic benchmark, with direct comparison to BC baselines. Diffusion policies model multimodal action distributions via conditional denoising; flow matching learns deterministic continuous-time transports enabling efficient single-step sampling. Targeted ablations on goal conditioning and observation horizon length isolated which design choices drive performance. Results showed diffusion policies achieve state-of-the-art on long-horizon manipulation, while flow matching provides a competitive and faster alternative.
Other Projects
Brain Tumor Classification using ResNet (January 2021 – April 2021)
Implementation and benchmarking of ResNet on TensorFlow for brain tumour scan classification. Evaluated multiple performance strategies (identity shortcut connections, batch normalization tuning) and deployed on a medical imaging dataset.
3-Band Audio Equalizer (March 2020 – May 2020)
Op-amp-based audio equalizer with independent gain control for Bass, Mid, and Treble bands, plus distortion detection and master gain control. Designed in MATLAB and simulated in Proteus.
Electronic Voting Machine (January 2020 – February 2020)
8051 microcontroller-based EVM with 16×2 LCD display for results and election status. Designed and simulated in Proteus.