K. Niranjan Kumar

I am a researcher with 7+ years experience in Reinforcement Learning, Robotics and Computer Vision. I completed my Ph.D. earlier this year at the Georgia Institute of Technology, where I was advised by Sehoon Ha and Irfan Essa. My research is in the union of

Robot learning + Legged robots + Interactive perception

I am in the job market for post-doc and industry research roles in robotics. Please reach out to me if you have any opportunities.

Email  /  CV  /  Google Scholar  /  Twitter

News

• 09/2024: Our work on Adaptive Bayesian Domain Randomization has been accepted to IROS 2024.
• 12/2023: I was invited to present my work at Can we build Baymax? workshop at Humanoids 2023.
• 11/2023: Presented my recent work "Words into action" at CoRL 2023 LangRob workshop.
• 08/2023: My work CCRL now published as an article on Robotics and Automation Letters 2023

Research

My current research focus involves creating learning-based policies that enable high-DOF robots to exhibit natural fluid behaviors, all while reducing the reliance on intricate reward engineering. To this end, my ongoing work involves using Large Language Models (LLMs) for iterative, human-guided refinement for learned control policies.

Learning diverse and expressive humanoid behaviors from human language instructions.

Tune simulation parameters to efficiently transfer to target environment.

Cascaded Compositional Residual Learning (CCRL), which learns composite skills by recursively leveraging a library of previously learned control policies.

Finding hidden objects in structured cluttered environments using deep reinforcement learning.

Learning interaction policies to estimate mass distribution of articulated objects in minimal steps.

Projects

Built an algorithm that converts an image into an infinitely looping video using a neural network and user input.

We generate realistic hand manipulations of object by using a Generative Adversarial Network by conditioning on the hand and finger pose. We adopt an encoder-decoder architecture where we input a rough rendering of the skeletal pose and get a realistic rendering of a hand as the output. We generate datasets with simulated and real human arms at different poses and use it to train our model. Given an hand pose and object pose we showcase a pipeline to generate an image conditioned on the poses.

Design an automated eye-dropper that detects the patients eye, adminsters the eye drop and registers if the adminstration was successful using CV (eye-tracking, blink detection, drop tracking).

Built a controller for tele-operated mobile robot that can be controlled using a computer keyboard.

Service

Teaching

Mentoring Experience

Template borrowed from here.