Mihir Parmar

Mihir Parmar

Graduate Student in Robotics

University of Pennsylvania

Hey, I’m Mihir!

Currently, a graduate student in the Robotics program of the GRASP Lab at Penn. An aspiring Roboticist, I am particularly excited about the potential impact of Robotics and Artificial Intelligence together in significantly improving the quality of our daily lives and I look forward to solve some of the most challenging problems in this area. My interests lie in the fields of planning, learning, vision and perception for Robotic systems.

Recently, under the guidance of Prof. Michael Posa, I have been working on leveraging the powerful inductive bias of Deep Learning to solve the challenging nature of frictional contacts. During my time at Penn so far, I have also worked on developing a full autonomy stack for F110 autonomous race cars and designing deep learning architectures for mainstream computer vision tasks such as classification, detection, tracking and segmentation.

Previously, I completed my Bachelor’s in Mechanical Engineering from National Institute of Technology Karnataka, Surathkal where I worked on motion-planning for multi-agent quadcopter systems and probabilistic robotic frameworks like Kalman Filters and SLAM. I have spent time as a Research Intern in the Robotics Lab at IIT Madras, advised by Prof. T. Asokan.

Always open to talk about anything related to Robotics, Soccer, Space Exploration and anything that requires out-of-the-box thinking!

🤖 🚀 ⚽ 🏕️ 🎮

Interests

  • Robotics
  • Computer Vision
  • Deep Learning

Education

  • MSE Robotics, 2021

    University of Pennsylvania

  • B.Tech in Mechanical Engineering, 2019

    National Institute of Technology Karnataka, Surathkal

Experience

 
 
 
 
 

Graduate Research Assistant

Dynamic Autonomy and Intelligent Robotics (DAIR) Lab, GRASP

May 2020 – Present Philadelphia
  • Develop Physics-inspired recurrent neural-network architectures on Python using PyTorch to learn stochastic, multi-modal transition dynamics in systems with contact.
  • Created simulation environment for data collection, testing and visualization in Mujoco, Bullet and Drake frameworks.
 
 
 
 
 

Graduate Teaching Assistant

University of Pennsylvania

Jan 2020 – May 2020 Philadelphia

Supported course development, conducted weekly office-hours, monitored and responded to student questions on communications for:

  • edX Robotics MicroMasters Program: Online program taught by GRASP faculty which includes 3 courses encompassing topics from various domains of Robotics and Learning
  • CIS581 Computer Vision: Class of 40 students taught by Prof. Jianbo Shi focusing on image processing concepts and deep learning for vision
  • CIS520 Machine Learning: Class of 120 students taught by Prof. Lyle Ungar focusing on mathematical foundations of machine learning.
 
 
 
 
 

Undergraduate Researcher

Robotics Lab, NITK

May 2017 – Jul 2017 Surathkal, India
  • Implemented a Centroidal Voronoi Configuration based search algorithm to achieve optimal deployment of multi-quadcopter system for maximizing the reduction in uncertainty density over the search space.
  • Simulated point agents on MATLAB and extended the simulation to AR Drones simulated in Gazebo.
  • Achieved reduction in uncertainty density below the threshold value of 0.1 in 8 search steps during simulation experiments.
 
 
 
 
 

Project Team Member

Daimler India Commercial Vehicles

Sep 2016 – Apr 2017 Surathkal, India

This project of developing a low-cost AGV was sponsored by Daimler India Commercial Vehicles.

  • Designed a novel optimised Material Handling System (MHS) for AGV.
  • Designed and implemented a servo-controlled toothed coupling between the trolley and AGV in an under-tugging position, reducing the torque and power required at drive motor in a 1:15 ratio while occupying minimum floor-space.

Video

 
 
 
 
 

Freelance Web Content Writer

Toppr

Jun 2016 – Aug 2018 Mumbai, India
Freelance content Writer for the online coaching platform toppr.com. Wrote informational articles primarily aimed at helping students preparing for engineering entrance examinations in India.

Projects

*
All Robotics Deep Learning Computer Vision

F1/10 Autonomous Racing

Developed a full-autonomy software stack for a F1/10 scaled autonomous racecar in ROS framework. Inclued autonomy modules- time-to-collision based automatic emergency braking, follow-the-gap reactive method, point-to-line metric based iterative-closest-point algorithm for scan-matching, A* and RRT* path-planner, April-tag based pose prediction, raceline optimization, trajectory tracking using Pure Pursuit and Model Predictive Control.

Hierarchical Controller for multi-agent Autonomous Racing

Implemented a hierarchical control strategy to compete in a virtual head-to-head racing of 1/10th scale autonomous cars. The controller employs a two-level structure- a high-level planner that generates a reference trajectory that maximizes the progress on the track ensuring obstacle avoidance. Next, we use MPC based tracker for tracking the reference trajectory obtained from the planner. The resulting strategy, capable of completing overtakes, successfully completed the race with a winning rate of 90%.
PDF Slides Video

Semantic Scene Composition for Amodal Instance Segmentation

A novel deep learning architecture capable of generating realistic occluded compositions by placing provided individual objects in the scene, in a context-aware manner while preserving the original full masks of the occluded object. Used a combined STN-GAN framework to learn a projection matrix based on encoded geometry and semantic information. Evaluated the generated compositions by fine-tuning a pre-trained Mask-RCNN for the task of amodal instance segmentation and reported the COCO-style mean-Average Precision (mAP) metric.
PDF Video

Particle Filter based RGB-D Tracking for Mobile Robots

A Particle Filter based approach for real-time object tracking using mobile robots with an RGB-D camera. The particle filter estimates the location of the object in the global frame and updates the weight of the particles by computing correlation using the 2D feature descriptors of the object inside the bounding box detection. The bounding box detection was obtained using the MobileNet-SSD detection framework.
PDF

Deep Learning for Vision

Implemented transfer learning on ResNet for image classification, and generative architectures:- Variational AutoEncoder, Least-Square GAN and DCGAN for image generation.

SLAM for THOR-OP humanoid robot

Implemented an Occupancy grid-based SLAM algorithm to generate the map of an indoor environment. Particle filters were used for localization of the robot, using measurements as scan from LIDAR, odometry data and the estimated map in the previous step.

Quaternion-based Unscented Kalman Filter for Orientation Tracking

Implemented an Unscented Kalman Filter to estimate the orientation of a rigid body using measurements from a 6 DOF IMU. UKF used the quaternion representation of the orientation to avoid singularity problems.

Sequential Models for Text Classification and Generation

Worked on data pre-processing of Amazon Reviews Dataset and implementing LSTM, BiLSTM, GRU, RNN architectures with Attention modules for review rating prediction using weighted loss and SMOTE techniques to handle class imbalance. Improved the F1 score accuracy by using b-directional transformer-based architectures BERT and RoBERTa. Designed seq2seq architecture with attention trained using teacher forcing strategy to generate summaries of review text.

Object Tracking using Optical Flow

Optical flow model to track multi-object moment in videos. Implemented Lucas-Kanade feature tracker and pyramidal KLT tracker to estimate object moments in successive video frames.
Video

Yelp Review Rating Prediction

Developed and implemented a CNN model for rating prediction and sentiment classification of YELP user restaurant-reviews. For the sentiment classification task, demonstrated that a simpler model which uses only adjectives of the review as its features yield similar results when compared to a complex model that utilizes the entire review.
PDF

Image Mosaicing

Created a Python API to stitch multiple images with overlapping field of views to create a panaromic effect employing Harris Corner Detector, Adaptive Non-Maximal Suppression, Histogram of Oriented Gradients feature descriptors, outlier rejection and homography estimation using RANSAC algorithm.

Contact