Gokul M K

I'm a Dual Degree student in Robotics at the Indian Institute of Technology Madras. I work with Prof. Anuj Tiwari at the DiRO Lab on Multi-agent Quadruped Robots. I also collaborate with Prof. Arvind Easwaran at the CPS Research Group, NTU Singapore, where I focus on trustworthy learning agents for F1TENTH Autonomous Racing.

I worked as an AI Research Intern at Qneuro India Pvt Ltd, developing deep learning models for inferring EEG, PCG Signals. I interned at HiRO Lab, IISC Bangalore, where I focused on learning for bimanual robot manipulation. I was part of Team Anveshak, the student-run Mars Rover team at IIT Madras that competes in international rover competitions. I initially contributed as an Embedded Engineer before being promoted to Electronics and Software Lead.

Email  /  CV  /  LinkedIn  /  Github

Research Interests

I explore Reinforcement Learning, Deep Learning, Generative AI, LLMs and Robotics. My research interests lie in developing intelligent, model-based, heirarchical controllers for robotic applications. In parallel, I explore advancing reasoning capabilities in LLMs and fine-tuning them using RLHF and SFT.

Improving the baseline tracking controller with residual Reinforcement Learning. Generalization across various racing tracks, static and dynamic obstacles.

Poster

Testing the MIT convex MPC framework for quadruped locomotion with various gaits. Addressing the limitations of the controller by data-driven Koopman method and learning the residual using Reinforcement Learning.

Developing an heirarchical learning-based controller for intelligent cooperative payload transport by a group of quadrupeds in varying terrains.

Implemented a Decentralized Training and Decentralized Execution (DTCE) approach for tackling swarm navigation in continously evolving environments.

Report | Code

Implemented the research by LASA Lab, EPFL on modulated dynamical systems for Coordinated Bimanual Robotic Manipulation. Also worked on implementing Tossingbot by learning the residual physics for throwing.

Compared various learning-based, analytical grasping algorithms and benchmarked their perfomance in simulation and hardware. Formulated a light-weight grasping algorithm using Euclidean clustering.

Video

A research-focused collection of from-scratch Transformer language models, exploring training stability, scaling behavior, and instruction fine-tuning. Currently, fine-tuning grootTiny (120 M) parameter model and pre-training grootSmall.

Code | Video

Addressing sub-optimality and diversity challenges in Offline-RL trained with large datasets collected for long-horizon tasks. Employed a heirarchical framework and validated its performance.

Report | Code

Optimized RIDNet, a deep learning model to denoise and deblur MVTEC dataset. It achieved a PSNR of 32.8 on the training examples and 34.78 on the test examples. Benchmarked the results with other architectures.

Report | Code

Compared two option learning methods, SMDP and Intra Option Q learning in the taxi gym environment. Intra-Option made faster updates while an action is taken rather than waiting for the option to end.

Report | Code

Continuously tracing a trajectory using RRT* while minimizing the deviation in the end-effector pose through an Optimal control formulation. This shares similarities with 3D printing task by mobile manipulators.

Report | Code | Videos

Programmed a Mobile Manipulator to identify, pluck and place coloured bell peppers inside a greenhouse.

Code | Video

Website source code from jonbarron