GVS Mothish

I am currently a First Year Phd Student at Computational and Data Sciences (CDS) in IISc, Bangalore, India, working on Generative models, Computer Vision, Deep Learning Architectures and Robot Learning under the supervision of Dr. Deepak Subramani. Prior to this, I completed my MTech Robotics and Autonomous systems at RBCCPS in IISc, Bangalore, India, where I focused on Reinforcement Learning and Generative models based locomotion of walking robots, under the guidance of Dr.Shishir.N.Y at Stochastic Robotics Lab at IISC. I am particularly interested in developing Intelligent Stochastic Systems.

My doctoral research focuses on Generative Models, Computer Vision, Neural Operators, Deep Learning.

Email  /  CV  /  Linkedin  /  Twitter  /  Github

Course Work @ IISc

Publications

* Denotes Equal Contributiion

Conditional Diffusion with Nonlinear Data Transformation Model (CN-Diff), a generative framework that employs novel nonlinear transformations and learnable conditions in the forward process for time series forecasting. A new loss formulation for training is proposed, along with a detailed derivation of both forward and reverse process.

TL;DR : A Diffusion Model based Walking Policies for Bipedal Robot Walking on Unseen Terrains

We have designed a real-time robot controller based on diffusion models, which not only captures multiple behaviours with different velocities in a single policy but also generalizes well for unseen terrains. Our controller learns with offline data, which is better than online learning in aspects like scalability, simplicity in training scheme etc.

TL;DR : An Inverse Reinforcement Learning based Target Prediction using Eye Gauge and Hand movement

A multimodal intent prediction algorithm involving hand and eye gaze using Bayesian fusion. Inverse reinforcement learning was leveraged to learn human preferences for the human-robot handover task.

TL;DR : A Modular Bipedal Robot design and Walking controller based on Linear Policies

This paper introduces the Stoch BiRo, a cost-effective bipedal robot designed with a modular mechanical structure having point feet to navigate uneven and unfamiliar terrains. The robot employs proprioceptive actuation in abduction, hips, and knees, leveraging a Raspberry Pi4 for control.

Articles Submitted for Peer Review - First Authored

Other Projects

• This was developed as a part of coursework for Robotic Perception course.

• PWC-NET a CNN-based approach was used to estimate the optical flow from the successive video frames.

• Yolo V5 was used for object detection in the given frame.

• Calculated the relative movement of predicted object’s movement with optical flow vectors

• Aim: Predict missing values in IoT sensor data (temperature and humidity) using Bayesian methods.

• Phases: Data analysis, pre-processing, and transitioning from Frequentist to Bayesian approaches.

• Bayesian Models: Bayesian Ridge Regression, Gaussian Process Regression, and PyMC3 used for accurate imputation.

• Object detection (Yolo) and Lane Detection (UltraFast Lane detector) module was developed using pytorch framework. While the Depth Estimation (MiDaS) was developed using TensorFlow framework.

• The inference time of the model was optimized using the Intel oneAPI Deep Neural Network Library on the intel’s developer’s cloud.

• A performance speedup of approximately 4.5x was achieved using the oneAPI libraries.

This is a Sponsored project under the supervision of Eyantra IIT-Bombay and MHRD, and a part of the Robotics Competition organized by Eyantra.

Awards and Acheivements

For the Outstanding academic performance in B.Tech

Theme: Breaking the wall of irrigation challenges with Machine Learning and Internet of Things

Worked with Intel AI Analytics Toolkits and Intel optimized frameworks such as TensorFlow, PyTorch.

Top up fellowship for the Masters project awarded by ARTPARK.