Atharva Mete

I am a 2nd-year MS student in the Georgia Tech School of Interactive Computing studying machine learning and robotics. I am advised by Prof. Animesh Garg and have been working on representation learning for robotic manipulation.

I graduated from IIT Bombay with a major in Aerospace Engineering and a double minor in 'Artificial Intelligence & Data Science' and 'System & Control Engineering.' I was advised by Prof. Shashi Ranjan Kumar and worked on cooperative multi-agent reinforcement learning problems.

I also led a student team that worked on autonomous drones!

CV (March 2023) / GitHub / Email / Twitter

Publications

QueST: Self-Supervised Skill Abstractions for Learning Continuous Control
Atharva Mete, Haotian Xue, Albert Wilcox, Yongxin Chen, Animesh Garg
(preprint - under review) [link]
A novel multitask and few-shot behavior cloning algorithm which first learns to tokenize continuous robot action sequences before using an autoregressive transformer to learn a policy in the token space.

Cooperative Multirobot Exploration using Reinforcement Learning. Atharva Mete, Malek Mouhoub
(Accepted to ICUAS 2023) [link]
Classical multi-agent exploration algorithms are quite inefficient owing to their independent nature of planning. How can we create intelligent agents capable of incorporating the high-level intentions of the other agents and long-horizon planning in exploration goal selection for a partially observable setting?

A Novel Approach to Autonomous Aerial Manipulation using Multiple Drones M.Agarwal, R.Shukla, K.Badoriya, R.Khantwal, H.Naik, M.Dixit, A.Mete, S.Sharma, A.Sonara, B.Prakash, D.Shukla.
(IARC Symposium on Dynamic Flight Behavior) [link]

Robotics Projects

International Aerial Robotics Competition

Objective: To autonomously perform aerial replacement of a 2kg heavy module attached to a mobile platform while operating over long distances.

Approach: Mother-daughter configuration, with the daughtership launched in-flight from the mothership. A Finite-State-Machine-based central autopilot was developed for decision-making and autonomy.
Module Detection: Implemented YOLOv4 on Jetson Xavier NX and optimized using TensorRT engines to speed up inference by 300%
In-flight Manipulation: Designed a custom robotic gripper and used a modified ROS Joint-State-Controller to perform the replacement.

Impact: Declared World Champions among 12 other top international universities

(GitHub) (Simulation Video)

Autonomous Indoor Drone -
Flipkart Grid Robotics Challenge

Objective: To develop a quadrotor capable of autonomously detecting and passing through frames in a GPS-denied environment

Localization: Deployed ORB-SLAM2 on Intel D435 RGBD camera and fused inertial & visual odometry data using Kalman filter to achieve pose estimation accuracy of 5cm

Frame Detection: Developed a robust cascade frame detection algorithm using Pointcloud & OpenCV with RGBD camera

Impact: Among the Top 3 National finalists, out of the 2000+ teams from other IITs and NITs

(GitHub) (Presentation)

Multi-Agent Collision Avoidance with Deep-RL

Introduction: The scalability of multi-agent approaches is bottlenecked by the connectivity requirements. How can we plan optimally in human-robot interaction scenarios where no state information is explicitly communicated?

The collision avoidance problem was formulated as a sequential decision-making problem, and Deep Q Learning was implemented to compute collision-free trajectories. Double DQN with Prioritized Experience Replay is used to accelerate learning and improve sample efficiency.

[Project GitHub Link]

Zero Episodes (supervised network)

500 Episodes

100 Episodes

1000 Episodes

Publications

Robotics Projects

International Aerial Robotics Competition

Autonomous Indoor Drone - Flipkart Grid Robotics Challenge

Multi-Agent Collision Avoidance with Deep-RL

Autonomous Indoor Drone -
Flipkart Grid Robotics Challenge