Kush Garg
I’m a Mathematics and Computing undergraduate at Delhi Technological University (DTU), working in applied robotics and autonomy. I have interned at General Autonomy as a Robotics Autonomy Engineer, developing semantic exploration, terrain-aware planning, and visual-inertial odometry for quadrupeds. I’ve also interned at MIT Media Lab on non-line-of-sight perception. With UAS-DTU, I placed 2nd at the DARPA Triage Challenge 2024 (USA) and 3rd globally at ICUAS in Greece.
My technical work spans autonomous exploration and navigation, SLAM, sensor fusion, visual-inertial odometry, deep learning (GANs, LaMa), physics simulation, and UAV systems and avionics.
I enjoy building and deploying products, robots, and drones, taking ideas from concept to real-world systems. Outside of work, you’ll usually find me flying drones or playing table tennis.
News
- 08/2026 Won 1st place nationally and a ₹6 lakh scholarship at the IET India Scholarship Award 2026.
- 07/2026 Completed a Robotics Autonomy Engineer internship at General Autonomy.
- 03/2026 Scaled RenderPerk to six-figure monthly revenue.
- 01/2026 Founded RenderPerk, a distributed compute rendering platform.
- 06/2025 Joined MIT Media Lab as a Research Intern.
- 05/2025 Paper accepted at ICUAS 2025 in Charlotte, NC.
- 10/2024 Won 2nd Place ($60,000) at the DARPA Triage Challenge with Team UAS-DTU.
- 10/2024 Obtained FAA Part 107 S-UAS Pilot License.
- 05/2024 Secured 3rd Position globally and 2nd in Simulation at the ICUAS UAV Competition in Greece.
Experience
Robotics Autonomy Engineer Intern, General Autonomy
More Info
Semantic Goal-Based Autonomous Exploration & Navigation for Quadrupeds
- Implemented and extended WildOS-based semantic exploration, enabling quadrupeds to discover frontiers, maintain navigation graphs, backtrack dead ends, and reach goal locations autonomously.
- Implemented semantic terrain-aware planning using Mask2Former, enabling physical quadrupeds to avoid rivers and dense bushes while preferring traversable regions.
Improving Visual-Inertial Odometry for Quadrupeds
- Improved factor-graph VIO to handle IMU impulses induced by quadruped foot-ground impacts.
- Improved robustness to jerky motion, dynamic foliage, moving shadows, and tall vegetation.
- Achieved 30% lower Z-drift and lower ATE than ZED SDK over 1 km+ outdoor and multi-floor indoor runs.
Founder & CEO, RenderPerk
More Info
- GPU rendering platform built on top of distributed compute with near-unlimited scale.
- Renders massive scenes across hundreds of GPUs globally in parallel.
- Up to 80% lower cost and 10x faster than traditional render farms.
- Used by 100+ animators, VFX studios, and artists worldwide, generating six-figure monthly revenue.
Research Intern, MIT Media Lab
More Info
- Developed SuperEx that enables robots to see beyond line-of-sight using SPAD-based NLOS sensing.
- Modeled a CUDA-accelerated real-time physics simulator for 360 multi-bounce photon scattering.
- Trained a GAN to reconstruct occluded geometry from noisy backprojection maps of NLOS data.
- Trained a LaMa-based inpainting model to predict complete indoor floor layouts from NLOS observations.
Both corridors appear identical in line of sight, but using NLOS, we can predict the dead end beforehand.
From a small observed map, NLOS lets us predict a much larger unseen map.
DARPA Triage Challenge
More Info
- Placed 2nd and won a $60,000 prize as a self-funded team in a U.S. Department of Defense competition.
- Engineered UAV and 3 UGVs integrating LiDAR, radar, cameras for autonomous casualty assessment.
- Designed multi-agent coordination strategies and developed a hybrid path-planning algorithm.
- Built a UAV geolocation pipeline with human detection achieving cm-level localization at 10 m altitude.
- Achieved autonomous, contactless assessment of real casualties using SLAM, path planning, and deep learning to estimate breath rate, heart rate, trauma, hemorrhage, and AVPU alertness from 1 m standoff.
Our developed systems, an NDAA-compliant drone and UGV.
DARPA real-world trials - rover triaging in the setting to assess casualties.
Path planning - autonomous routing and coordination for casualty assessment.
ICUAS UAV Competition
More Info
- Secured 3rd position globally and 2nd in simulation for developing an autonomous fruit-counting UAV.
- Implemented LiDAR–camera fusion for fruit counting and mapping in ROS2 with a hybrid A* TSP planner.
- Achieved 98% accuracy in simulation & 90% accuracy in real-world implementation.
- Stack: ROS2, Gazebo, LiDAR-camera fusion, mapping, path planning, and clustering.
- Sensors: Intel D455 and Velodyne VLP-16.
Our team
Camera-LiDAR fusion
Lidar Mapping
Fruit tracking and counting
Avionics Engineer, Team UAS-DTU
More Info
- Built autonomous UAV systems implementing SLAM, sensor fusion, RL and onboard control stacks.
- Executed 100+ hours of automated missions with Mission Planner and SITL-Gazebo workflows.
- Designed and fabricated quadcopters, VTOLs, fixed-wings, and rovers with comm systems.
- Developed ROS pipelines and deployed RealSense D455 for autonomous navigation and mapping.
Awards & Recognition
IET India Scholarship Award 2026
Aug 2026
National Winner | ₹6 Lakh Scholarship
- Won 1st place nationally among 48,000+ applicants and received a ₹6 lakh scholarship for the Context-Aware Assistive Wearable for Neurodivergent Children.
FAA Part 107 S-UAS Pilot License
Oct 2024 - Present
Federal Aviation Administration, USA
- Licensed to operate small unmanned aircraft systems in the United States.
Publications
SuperEx: Enhancing Indoor Mapping and Exploration using Non-Line-of-Sight Perception
Preprint, Under Review (2025)
More Info
- A framework for indoor mapping and exploration that uses non-line-of-sight sensing.
- Uses single-photon LiDAR to carve empty NLOS regions and reconstruct nlos region.
- Combines Pix2Pix and LaMa with global map prediction to improve frontier exploration.
- Reported a 12% gain in mapping accuracy under 30% coverage on simulated maps.
Projects
Context-Aware Assistive Wearable for Neurodivergent ChildrenESP32-S3, E-paper Display, Speech Recognition, Language Models
Context-Aware Assistive Wearable for Neurodivergent Children
ESP32-S3, E-paper Display, Speech Recognition, Language Models
- Built a small wearable that helps neurodivergent children communicate in everyday situations without needing to speak or navigate a complex phone screen.
- With one button press, it listens to the conversation, understands the immediate context, and shows a short suggested response on a calm e-paper display.
- It can also make traffic noise less distracting through wired earphones, while keeping nearby speech understandable.
- Designed privacy-first controls that keep recent conversations separate from confirmed personal preferences, so the wearer stays in control.
Working prototype with a quiet e-paper display and wired earphones.
Example flow: recognize a need, reduce distracting traffic noise, receive a suggested response, and complete the order.
Physics-Informed Neural Networks for Singular Perturbation ProblemsPyTorch, PINNs, ODEs
Physics-Informed Neural Networks for Singular Perturbation Problems
PyTorch, PINNs, ODEs
εu''(x) + (2x + 1)u'(x) + 2u(x) = 0, x ∈ [0,1]
with boundary conditions u(0) = 0 and u(1) = 1.
- Developed an adaptive moving-mesh PINN architecture to solve singularly perturbed ODEs with sharp boundary layers and stiff solution regimes.
Aerial Human Detection and Geo-localizationQuadcopter, PiCam, GPS, IMU, Python
Aerial Human Detection and Geo-localization
Quadcopter, PiCam, GPS, IMU, Python
- Implemented PiCam-based human detection with intrinsic–extrinsic calibration and IMU correction.
- Achieved cm-level geo-localization by fusing GPS, camera transforms, and UAV state estimates.
Drone On Time – Autonomous Drone Delivery SystemIncubated at DTU Innovation and Incubation Foundation
Drone On Time – Autonomous Drone Delivery System
Incubated at DTU Innovation and Incubation Foundation
- Built an autonomous medical drone delivery system achieving 75% cost and 70% emission reduction.
- Completed 25+ successful autonomous deliveries carrying upto 2 kg payloads over a 10 km range.
This video shows an autonomous payload drop from a UAV at 10 meters altitude.
Dynamic Multi-UAV Obstacle AvoidanceROS2, Gazebo, MAVROS, Ardupilot, Python
Dynamic Multi-UAV Obstacle Avoidance
ROS2, Gazebo, MAVROS, Ardupilot, Python
- Built a ROS2–Gazebo obstacle avoidance pipeline using Velodyne 360° LiDAR point clouds.
- Enabled 15 m/s collision-free flight via MAVROS–ArduPilot integration and dynamic avoidance.
- Stack: ROS2, Gazebo, MAVROS, and APFA.
This video demonstrates dynamic obstacle avoidance for a UAV in flight.
Skills & Platforms
Languages & Tools
Python, C/C++, MATLAB, PyTorch, JAX, NumPy, CUDA, Numba
Estimation & Planning
VIO, Factor Graphs, MPPI, Autonomous Exploration, Frontier Planning, SLAM, Sensor Fusion, Camera Calibration
Perception & Sensors
Velodyne LiDAR, Livox LiDAR, mmWave Radar, RealSense D455, ZED Camera
Hardware & Embedded
Jetson Nano/Orin, Raspberry Pi, ESP32, Pixhawk
Robotic Platforms
Quadrupeds, Humanoids, Quadcopters, Fixed-Wing UAVs, VTOLs, Avionics
Simulation & Communication
ROS1/2, Gazebo, MuJoCo, Unreal Engine, ArduPilot, SITL, gRPC, ZMQ
Education
Delhi Technological University (DTU)
Aug 2023 - Present
Bachelor of Technology in Mathematics and Computer Science | CGPA: 8.57 / 10
Delhi, India
More Info
- Computer Science: Artificial Intelligence, Machine Learning, Computer Vision, Data Structures & Algorithms, Operating Systems, Computer Networks, Theory of Computation.
- Mathematics: Linear Algebra, Modern Algebra, Differential Equations, Probability & Statistics, Stochastic Processes, Discrete Structures, Complex Analysis, MATLAB.
- Academics Head, Mathematics and Computing Community, DTU.
Modern School Barakhamba Road
April 2021 - June 2023
Senior Secondary Education (CBSE Class XII) : 94.4%
Delhi, India
More Info
- Vice President, Tech Society, Modern School Barakhamba Road.
Gallery