Srinath Mahankali

I am an undergraduate student at MIT where I am advised by Pulkit Agrawal as part of the Computer Science and Artificial Intelligence Laboratory (CSAIL). I am very interested in deep learning, reinforcement learning, and their application to robotics. Currently, I am working on improving exploration in deep reinforcement learning and developing energy-efficient controllers for quadruped locomotion through sim-to-real reinforcement learning.

Previously, I worked on inverse problems with Yunan Yang and studied the linear separability of one-layer neural networks with Promit Ghosal.

Updates

Research

We improve exploration in both discrete and continuous control domains by optimizing random reward functions parameterized by a sampled latent vector.

We train energy-efficient policies for quadruped locomotion tasks while improving task performance through constrained reinforcement learning.

Randomly generated rewards can explain a significant fraction of exploration improvements from novelty-based intrinsic motivation.

We prove bounds on the convergence and stability of the inverse scattering series under different Sobolev norms, finding conditions under which the radius of convergence and stability are improved.

Randomly initialized one-layer neural networks, with high probability, make datasets linearly separable.

Full waveform inversion with an optimal transport-based objective has superior convexity compared to the standard least-squares objective function for certain velocity models.