Ankur Deka

I am a Robotics ML Engineer at Tesla working on the manipulation skills of Optimus.

Prior to this, I was an AI Research Engineer/Scientist working on Reinforcement Learning for robot navigation and manipulation at Intel Labs in Matthias Mueller's group. I was an AI Robotics Engineer at Nimble Robotics developing state-of-the-art AI technologies for automating robotic warehouses.

I did my Master's at the Robotics Institute, Carnegie Mellon University (CMU) working on Multi Agent Reinforcement Learning (MARL) and Imitation Learning advised by Prof. Katia Sycara and Prof. Michael Lewis. My Master's thesis is On Combining Reinforcement Learning and Adversarial Training.

I completed my BTech in Electronics and Commmunication Engineering and Computer Science from Indian Institute of Technology (IIT) Guwahati in 2019. I have worked on Socially-Aware robot navigation at ATR IRC Lab, Japan in Dr. Norihiro Hagita's group and on 3D reconstruction and segmentation of medical images at Chubu University, Japan in Prof. Yuji Iwahori's group.

I have received the J N Tata Scholarship for graduate studies and was selected for KVPY (English: Young Scientist Incentive Plan).

Email  /  Google Scholar  /  LinkedIn  /  GitHub  /  Resume

• [May 2024] OpenBot-Fleet presented at ICRA 2024.
• [May 2024] OpenBot-Fleet accepted to ICRA 2024.
• [July 2023] I joined Optimus (Tesla's Humanoid Robot project) as a Robotics ML Engineer.
• [Mar 2023] I joined Nimble Robotics as an AI Robotics Engineer.
• [Jan 2023] Our paper got accepted to ICRA 2023.
• [Sept 2022] I presented our paper at CoRL 2022.
• [Sept 2022] Our paper got accepted to CoRL 2022.
• [Sept 2021] I presented our paper at IROS 2021.
• [Sept 2021] I joined Intel Labs as an AI Research Engineer/Scientist.
• [Aug 2021] I graduated from CMU Robotics Institute.
• [July 2021] I presented our paper at ICSI 2021.
• [June 2021] Our paper has been accepted to IROS 2021.
• [March 2021] Our paper has been accepted to HFES 2021.
• [Feb 2021] Invited to deliver a talk on Reinforcement Learning by AAII.
• [Sept 2020] Invited to talk at RedOne Technologies, South Korea about my work on MARL.
• [July 2020] Presented our paper in a Spotlight Talk at RSS 2020 MRTA Workshop.
• [Sept 2019] Joined CMU RI as graduate student.
• [July 2019] Received J N Tata scholarship for graduate studies.
• [June 2019] Graduated from IIT Guwahati.
• [Jan 2018] Presented our paper and received Best Poster Award at Bioimaging 2018 Conference, Portugal.
• [Jan 2018] Started an the initiative IgnoTech.
• [Oct 2016] Received 1st Position in Robothlon 2016.
• [Aug 2015] Joined IIT Guwahati as an undergraduate student.
• [May 2013] Received Brainstorm scholarship.

I am interested in Artificial Intelligence (AI) agents that can act in the world, learn from their experiences and adapt to changes in the environment. Such versatile agents would ultimately improve the quality of human life - robots would provide relief in disasters, perform daily chores in homes and help perform surgeries. I have explored areas of Reinforcement Learning, Computer Vision, Deep Learning, Probabilistic Modeling and hardware design of mechanisms/electronics.

A new class of RL algorithms that improves Adversarial Imitation Learning by leveraging reward differentiability.

Reinforcement Learning policies trained purely in simulation perform compliant insertion of dishes and cups with noisy observations by leveraging a soft gripper and force feedback.

MARL can generate complex leader-follower multi-robot navigation strategies that hide the leader's identity for defense against enemies.

Complex multi-agent strategies can naturally emerge out of competition between teams.

Learning to detect nuclei boundaries can help improve segmentation.

Proposed an Spatio-Temporal Graph based framework for pedestrian trajectory prediction along with an improved attention model.

Reconstruction of 3D shape of polyp from 2 monocular endoscopy images with unknown transformation.

Sim2Real transfer of low-level control policy for self-driving and drifting using q-learning. The project involved creating a simulation environment and training a policy in simulation, modifying the hardware of an RC car, setting up visual trackers, integrating the system with Robot Operating System (ROS) and finally transferring the simulation policy to the real world.

A ball robot similar to BB8 Droid, that can be wirelessly controlled from a long distance to maneuver. There are no external actuators and the robot looks like a normal ball from the outside.

I was bad at the Chrome browser's dinosaur game, so I made a robot to play it.

I made an image scanner from a broken CD drive.

I made made my first robot, a remote-controller pick and place machine from broken toys.

Thanks to Jon Barron for providing the nice website template website template.