Soumith Udatha

I am a second year Masters' Research Student at Carnegie Mellon University, where I work on bridging the gaps for applying Reinforcement Learning to the systems in the real world. I finished my undergraduate studies in Mechanical Engineering from the Indian Institute of Technology, Bombay(IIT Bombay).

Email / CV / LinkedIn

Research

I'm interested in implementing Reinforcement Learning agents for applications in the real world. Making RL agents use data and representations in the world effectively is a higher level objective. Some of the applications I am currently looking at include RL agents for Autonomous Driving Scenarios, Goal-Conditioned RL for D4RL environments and CARLA Autonomous Driving challenges and in building co-operative agents for Human-Agent Collaboration tasks.

Updates

"Reinforcement Learning with Probabilistically Safe Control Barrier Functions for Ramp Merging" got accepted to ICRA 2023
"Reinforcement Learning with Probabilistically Safe Control Barrier Functions for Ramp Merging" is in the TechExplore Article now and other Media.

Imitating Past Successes can be Very Suboptimal
Ben Eysenbach, Soumith Udatha, Sergey Levine, Ruslan Salakhutdinov,
36th Conference on Neural Information Processing Systems (NeurIPS 2022), New Orleans, USA
arXiv / code / video

One seemingly-simple way of doing RL is to do imitation learning on successful trajectories, and prior methods like goal-conditioned imitation learning use this to great effect. For draw a connection between these prior methods and reward maximization, showing that these prior methods do not quite correspond to reward maximization, and actually get be worse than doing nothing. Our analysis suggests a simple fix.

Safe Reinforcement Learning for Ramp Merging with Probabilistically Safe Control Barrier Functions
Soumith Udatha, Yiwei Lyu, John Dolan
39th International Conference on Machine Learning (ICML 2022),Baltimore, USA (Workshop)
2023 IEEE International Conference on Robotics and Automation (ICRA 2023) (Conference)
arxiv /

Prior work has looked at applying reinforcement learning and imitation learning approaches to autonomous driving scenarios, but either the safety or the efficiency of the algorithm is compromised.With the use of control barrier functions embedded into the reinforcement learning policy, we arrive at safe policies to optimize the performance of the autonomous driving vehicle.

A novel foot interface versus voice for controlling a robotic endoscope holder
Yanjun Yang Soumith Udatha, Dana Kulic, Elahe Abdi,
2020 8th IEEE RAS/EMBS International Conference for Biomedical Robotics and Biomechatronics (BioRob)
paper /

An ergonomocal and intuitive foot interface for assisted robotic surgeries. We propose a novel foot interface to give the surgeon direct control over a robotic camera holder to replace the human camera holding assistant. The foot interface should allow the surgeon to control the camera while their hands are occupied with the primary surgical task. It can control 4 degrees of freedom of the camera’s pose at 2 speeds.

Want to make such a website? source code,