Biography
Hi! I am Karmesh, an AI Resident at FAIR, working with the Habitat team under the supervision of Dr. Oleksandr Maksymets and Prof. Dhruv Batra.
I’m interested in robot learning, and have been researching how can we learn better visual representations for downstream training of Embodied AI agents. Previously, I worked as a Senior Robotics Engineer at ISEE, an autonomous vehicles startup based out of Boston, trying to automate the yard trucking space. I completed my Masters in Robotics Systems Development (MRSD) at the CMU Robotics Institute in 2020.
Download my resumé.
- Reinforcement Learning
- Robot Learning
- Meta Learning
Masters in Robotic Systems Development, 2020
Carnegie Mellon University
B.Tech in Mechanical Engineering, 2017
Indian Institute of Technology, Guwahati
Experience
- Researching self-supervised pretraining techniques for learning useful representations for embodied agents.
- Using the learnt representations on downstream RL tasks like image-goal navigation and object-goal navigation in the Habitat Simulator.
- Explored deep uncertainty estimation techniques for predicting the closed loop tracking performance of an autonomous vehicle controller. Estimated the collision probability of the AV with respect to obstacles in an occupancy grid.
- Improved the trajectory optimization planner and robustified its collision checking. This led to an increased confidence in its performance and resulted in its deployment on the AV.
- Developed the speed planning module for safely achieving three-fold increase in the operating speed of the AV.
- Built toolboxes to automate the system identification and calibration procedure of Isee’s vehicles.
- Researched and implemented various vehicle and tire models for control application in AVs.
- Created a Q-Learning based planner to prevent monocular slam failure on non-holonomic robots.
- Developed a simulation environment in Gym-Gazebo for training, with Navigation Stack for planning and ORB-SLAM for perception and localization.
- Optimized ORB-SLAM and made it more robust to fuse its position output with RTK-GPS, IMU & Wheel Encoder data using an EKF.
- Worked on SLAM pose covariance estimation and extrinsic calibration of IMU and cameras
- Developed a two-level motion planner, utilizing the A-star (A*) and Rapidly-exploring Random Tree (RRT) algorithm, on a local map built using laser scanners for an electric golf cart.
- Created a vehicle model and forward-simulated the vehicle trajectory using Pure Pursuit steering controller and Proportional-Integral (PI) speed controller.
Recent Posts
Publications & Preprint
Projects
Soft Robotic Hand
Learning Dexterous Manipulation with a Low-Cost Soft Robotic Manipulator
CiLQR for Motion Planning
Constrained iLQR for Motion Planning in Autonomous Vehicles
Collision Avoidance System
Detection and Response for Potential Head-On Vehicle Crashes
Autonomous Driving using IRL
Learning from Demonstrations with Inverse Reinforcement in CARLA
Very Social Robot
Learning Human-Robot Tool Handover using Interaction
Formula Bharat
A formula-styled race car developed in IIT Guwahati
Recent & Upcoming Talks
Contact
- ykarmesh@gmail.com
- Santa Clara, CA 95054
- Monday 17:30 to 19:00 EST
Wednesday 17:30 to 19:00 EST - Book an appointment
- Tweet