Perception Engineer II at MVP Robotics · open to senior perception & CV roles

Aadesh Varude

Perception & Computer Vision Engineer

I build real-time perception that survives the real world — deep-learning obstacle detection, SLAM and multi-sensor fusion for autonomous mobile robots operating off-road, day and night, on compute that fits on the robot.

Currently shipping the perception and state-estimation stack at MVP Robotics on NVIDIA Jetson AGX Orin. MS Robotics Engineering, WPI (4.0 GPA).

Vermont, USA Authorized to work in the US 2+ years in production robotics
Portrait of Aadesh Varude

Research Interests

  • Computer Vision & Perception
  • Simultaneous Localization & Mapping
  • Sensor Fusion & State Estimation
  • Deep Learning
  • Classical Controls
92%
Obstacle detection accuracy across multi-terrain, day & night
−25%
Training data required vs. RGB-based baselines
2×
Faster semantic NeRF training (36h → 18h), no quality loss
4.0
GPA, MS Robotics Engineering at WPI
What I do

Perception, end to end

From sensor driver to deployed inference loop. I work across the full perception stack — and I care most about the part where research meets a robot that has to keep running.

Vision & Deep Learning

Real-time detection, segmentation and tracking. Novel image-encoding methods that cut labelled-data needs while holding accuracy in degraded, low-light and off-road scenes.

SLAM & State Estimation

Mapping and localization for mobile robots, plus visual-inertial odometry. MSCKF-style filters, pose graphs, and keeping drift bounded when GNSS drops out.

Multi-Sensor Fusion

LiDAR, stereo & mono cameras, GNSS and IMU fused into one consistent world model — with calibration, time sync and graceful degradation when a sensor lies.

Embedded Deployment

Getting models off the workstation and onto the robot. CUDA and tinyCUDA optimization, NVIDIA Jetson AGX Orin and Google Coral, ROS/ROS2 integration, Docker.

Experience

Where I've built things

Production autonomy, industrial research, and academic robotics labs.

Perception Engineer II

MVP Robotics
May 2024 — Present
Vermont, USA
  • Designed and implemented a real-time obstacle detection neural network with a novel image encoding method for autonomous mobile robots in multi-terrain, day and night outdoor environments — 92% detection accuracy while reducing training data requirements by 25% versus RGB-based methods.
  • Developed SLAM and a hybrid Vector Field Histogram (VFH+) path planner for real-time obstacle avoidance running on NVIDIA Jetson AGX Orin.
  • Built a human detection and tracking pipeline integrated into the robot's production autonomy kit.
  • Developed the autonomy perception stack for SAGE, a modular robot compute platform, supporting LiDAR and camera sensor integration.
  • Engineering a GNSS + visual-inertial odometry (VIO) sensor fusion stack to improve real-time positioning accuracy in GPS-denied environments.
PyTorchC++ROS2 CUDAJetson AGX OrinSLAM VIOLiDARGNSS/IMU

Research Intern — Computer Vision & Deep Learning

Nokia Bell Labs
Jun 2023 — Aug 2023
New Jersey, USA
  • Automated semantic label generation for NeRF training using Grounding DINO and Segment Anything (SAM), eliminating manual annotation entirely.
  • Optimized semantic NeRF training with a Tiny CUDA-based implementation, cutting training time by 50% (36 → 18 hours) with no loss in rendering quality.
NeRFtinyCUDAGrounding DINO SAMNerfactoPyTorch

Graduate Researcher

Adaptive & Intelligent Robotics Lab (AIR Lab), WPI
Aug 2023 — Dec 2023
Massachusetts, USA
  • Developed an obstacle avoidance stack for an autonomous underground cable-detection robot.
  • Designed a human–robot interaction interface providing real-time camera feedback, live map updates, initial navigation and teleoperation for a cable-following robot.
ROSQt/QMLNavigationTeleoperation

Earlier

Summer Research Intern Technische Universität Hamburg (HULK humanoid team) — walking trajectory optimization and vision for RoboCup 2022 · Hamburg, DE
Lab Member & Treasurer IvLabs, VNIT — reconfigurable snake robots, gesture control, published research 2019–2022 · Nagpur, IN
Selected work

Projects

Implementations spanning classical geometry, learned perception, state estimation and planning. Click any project for the full write-up, demo videos and results.

3D Vision

Reconstructed 3D scenes and recovered monocular camera poses with the classical pipeline — feature matching, triangulation, PnP, bundle adjustment — then rebuilt the same scenes with a tiny NeRF.

Python · OpenCV · PyTorch View details
3 videos

Perception

Horizon and field-edge detection via HSV masking, Canny and Hough transforms, followed by YOLOv4 ball detection on live RoboCup football footage.

Python · OpenCV · YOLOv4 View details

Deep Learning

Embedded standard and modified attention blocks into U²-Net for salient object detection, comparing attention-inside, attention-outside and combined architectures.

PyTorch · Python View details

Estimation

Multi-State Constraint Kalman Filter fusing IMU and stereo camera measurements for drone odometry, with a sliding window of camera pose states.

Python · NumPy View details

Geometry

From-scratch implementation of Zhang's method for intrinsics, extrinsics and radial distortion, with non-linear refinement of the reprojection error.

Python · OpenCV · SciPy View details

Classical CV

Probability-of-boundary detection combining oriented filter banks with texture, brightness and colour gradients — markedly cleaner boundaries than Canny or Sobel.

Python · OpenCV View details
Demo

Classical CV

Seamless panoramas from overlapping images — corner detection, ANMS, feature descriptors, RANSAC homography estimation and blending.

Python · OpenCV View details
Demo

Planning

A heuristic A*-based MAPF approach using feasible motion primitives that respect non-holonomic constraints, so plans are directly executable by real vehicles.

C++ · Search & Planning View details

Planning

Reference implementations of PRM, RRT, RRT* and Informed RRT*, alongside classical graph search (A*, Dijkstra, BFS, DFS) with comparative benchmarking.

Python View details
Video

Robotics

A reconfigurable robot that transforms between snake, quadruped and biped modes to cross terrain no single morphology handles well. Published at IEEE CASE 2021.

Robotics · Mechanism Design View details
Demo

Controls

Trajectory planning and generation for humanoid walking and kicking at TUHH, plus a PID controller for balance stabilization. Built for RoboCup.

Python · MATLAB · Controls View details
Video

Embedded

Controlling laptop functions through hand gestures sensed by ultrasonic sensors and a microcontroller — no camera, no wearables. Published research.

Embedded C · Ultrasonic Sensors View details
Demo

RL

Trained agents with DQN and PPO variants across Atari environments — replay buffers, target networks, advantage estimation and the usual stability pitfalls.

PyTorch · Gym View details

More on github.com/AadeshVarude — face swapping, SCARA kinematics and control, and coursework repositories.

Research

Publications

Peer-reviewed work in robotics and embedded systems.

IEEE CASE 2021 · Conference Paper

ReQuBiS — Reconfigurable Quadrupedal-Bipedal Snake Robots

H. Zade, A. Varude, K. Pandya, A. Kamat, S. Chiddarwar, R. Thakker

17th IEEE International Conference on Automation Science and Engineering (CASE), 2021, pp. 2241–2246.
doi: 10.1109/CASE49439.2021.9551526

Additional publications

  • Hand Gesture Control of Computer Features — R. Runwal, S. Dhonde, J. Pardhi, S. Kumar, A. Varude, et al.
  • Evolution of Hardware Trojans: Structure, Taxonomy, Countermeasures and Challenges — B. Jogi, A. Varude, G. Deshmukh, A. Bhurane, A. Kothari
  • Adiabatic Logic Gates using Wireless Charging — B. Jogi, A. Varude, A. Bhurane, A. Kothari
Toolbox

Technical skills

What I reach for day to day.

Languages

PythonC++C

Vision & Deep Learning

OpenCVPyTorchTensorFlow CUDAtinyCUDANerfacto Grounding DINOSegment Anything

Robotics & Simulation

ROSROS2Gazebo MATLABV-REPSLAM VIOSensor Fusion

Embedded & Sensors

Jetson AGX OrinGoogle CoralLiDAR GNSS/IMUStereo CamerasMono Cameras

Tools & Platforms

GitDockerLinux / Ubuntu Qt / QMLLaTeX
Education

Academic background

Graduate robotics at WPI, built on an electronics and communication engineering foundation.

MS, Robotics Engineering
Worcester Polytechnic Institute
Worcester, Massachusetts · Aug 2022 – May 2024
GPA 4.0 / 4.0
B.Tech, Electronics & Communication Engineering
Visvesvaraya National Institute of Technology
Nagpur, India · Jul 2018 – May 2022
GPA 8.45 / 10IvLabs robotics lab

Selected graduate coursework

RBE 502Robot ControlWPI · Aug–Dec 2022
CS 549Reinforcement LearningWPI · Aug–Dec 2022
CS 541Deep LearningWPI · Jan–May 2023
RBE 550Motion PlanningWPI · Jan–May 2023
RBE 549Computer VisionWPI · Aug–Dec 2023
RBE 595Visual–Inertial OdometryWPI · Jan–May 2024
Get in touch

Let's talk perception

I'm open to senior perception and computer vision roles, and always happy to talk shop about SLAM, sensor fusion, or getting models to run fast on embedded hardware.