Suveer Garg
Robotics Software Engineer, University of Pennsylvania
I am a full-stack roboticist interested in building autonomous systems that leave the laboratory and solve real-world problems. I graduated in May 2021 with a Masters in Systems Engineering from the University of Pennsylvania. As a senior robotics software engineer at Oishii, I maintain the operations of 60 production strawberry harvesting robots with minimal downtime. My work resulted in improving pick accuracy by 25% and reducing collateral damage to under 5%. During my two years at the Samsung AI Center in NYC, I contributed to research in computer vision, path planning and controls. The research resulted in 4 publications in top robotics conferences(ICRA and IROS) and 3 patents. I also engineered full autonomous systems on a mobile base and 7-DOF manipulator to demonstrate the applications of our research. I am looking for full-time roles in computer vision, path planning or machine learning.
- Python, C++, C
- Sklearn, Numpy, tensorflow, pandas, Matplotlib, PyTorch, MLFlow, Sagemaker
- Altium Designer, Atmel Studio, Visual Studio, Docker, k8s, Django, PostgreSQL
- Linux, ROS, AWS
- Quick Prototyping Boards (Arduino, Raspberrry Pi), SAMD21(ARM 32 bit)
Robotic harvesting system for Strawberries
Maintained 60 robots in production with minimal downtime. Improved pick accuracy by 25% and reduced collateral damage to under 5% by using 3D vision solutions in production. Created an ML training pipeline to improve mAP values of instance segmentation and keypoint detection R-CNN models. Developed system to collect pick-metrics (successful, missed & dropped), vision-system metrics (latency & rejections) and farm-metrics (coverage & berry counts), helping multiple teams make data-driven decisions.
HIO-SDF - Hierarchical Incremental Online Signed Distance Field
Developed a novel method for constructing a differentiable global SDF for real-time planning in unknown environments. HIO-SDF employs a coarse voxel grid that captures the observed parts of the environment together with high-resolution local information to train a neural network that represents the environment as a Signed Distance Field (SDF). More details in the project page linked below.
https://samsunglabs.github.io/HIO-SDF-project-page/https://arxiv.org/pdf/2310.09463.pdf
https://github.com/SamsungLabs/HIO-SDF
RAMP - Hierarchical Reactive Motion Planning for Manipulation Tasks Using Implicit Signed Distance Functions
Reactive Action and Motion Planner (RAMP), a hierarchical approach where a novel variant of a Model Predictive Path Integral (MPPI) controller is used to generate trajectories which are then followed asynchronously by a local vector field controller. RAMP can rapidly find paths in the robot's configuration space, satisfy task and robot-specific constraints, and provide safety by reacting to static or dynamically moving obstacles. More details in the project page linked below.
https://samsunglabs.github.io/RAMP-project-page/https://arxiv.org/pdf/2305.10534.pdf
https://github.com/SamsungLabs/RAMP
Path planning, minimum jerk trajectory generation & PID control for quadrotor
Path planning using A*, smoothing using piecewise minimum trajectory generation and PID control for quadrotor navigation in an indoor obstacle course. Additionally implemented a VIO pipeline for state estimation.
Deep Learning for Computer Vision
Implemented YOLO, SOLO, VAE, DCGAN, Cyclic GAN and Faster Mask-RNN. Course Project - Experimented with a state of the art method that uses Ordinal depth supervision for 3D human pose estimation from single 2D images. This method tackles the lack of availability/variability of 3D ground truth pose data for natural images. Please visit the github links below for more infomation on these projects.
https://github.com/suveergarg/SOLOhttps://github.com/suveergarg/YOLO
https://github.com/suveergarg/Faster-RCNN
https://github.com/suveergarg/VAE-GAN
https://github.com/suveergarg/PoseReconstruction
Pick and Place using the Lynx arm in Gazebo
Used Forward and Inverse velocity and position kinematics to stack blocks in a simulation environment. Won a stacking competition as part of the class final project.
RRT*, Particle Filter Based SLAM and Autonomous Quadrotor Landing on Moving Platform
Implemented Unscented Kalman Filter, particle filter SLAM with LIDAR data, RRT* algorithm. Simulated autonomous quad-rotor landing on a moving platform using Aruco markers and minimum jerk quad-rotor trajectory using Gazebo and ROS.
https://github.com/suveergarg/uav_landingImage Stitching and Optical Tracking
Created python modules for edge and corner detection, adaptive non-maximal suppression, homography estimation, feature matching using knn, outlier detection with RANSAC and poisson blending to produce smooth stitched images and videos. Additionally, implemented Lucas Kanade optical flow tracking algorithm to track moving objects in a video.
https://github.com/suveergarg/cis581-3AIntelligent Travel Friendly Medicine Dispenser
Designed an IoT medicine dispenser with automatic alerts for refilling, dose collection and location tracking. The end-to-end design process involved part selection, PCB layout, CAD design, power architecture design, Over The Air Firmware Update and custom bootloader design, wireless communications and firmware design.
https://github.com/briankwn/SmartDoseMulti-Robot Parcel Sorting System
Designed a ROS based simulation for multi-robot parcel sorting system. Developed software for map creation, planning, movement, diagnostics and evaluation of 120 robots. Introduced improvements which quadrupled the throughput of the system.
Ecofrost Technologies, Internship, Pune 2017-2018
Built capacitive sensing based embedded prototype for charge level measurement in Thermal batteries using level sensing of phase change material for energy management and temperature control in solar powered cold-storage units. Rectified errors after analysing data collected from prototype. At the end of the internship, I presented my results and offered a new approach to attain the objective of energy management for the system using the same sensor based on the learnings from the test experiments.
Smart Dustbin Project, WeConvert, New Delhi 2017-2018
The Smart dustbin was a Raspberry-Pi based system to classify waste objects and reward the user with credits based on what is dropped in the bin. It makes use of Image Processing algorithms to classify objects and sort trash. My work involved developing the detection algorithms, human machine interface and firmware for the embedded systems.The system was tested at various colleges, malls, and public places. The initiative was covered by media agencies such as DuBeat, Your Story and ZeeTv.