Shreyas Bhat

Shreyas Bhat

Human-Robot Interaction, Trust in HRI, Human-Behavior Modeling

Projects

Trust-Driven Human-Robot Interaction

In this project, we study how trust evolves when humans repeatedly interact with a robot recommendation system. We try to leverage quantitative models of trust to predict human behavior and design interaction strategies for the robot to promote trust in the robot’s recommendations.

We model the interaction as a trust-aware Markov Decision Process (trust-aware MDP) that consists of States, Actions, Transition Function, Reward Function, and Human Behavior Model… Read more

Trustworthy Human-AGV Interaction

In this project, we try to generate trustworthy behaviors for Automated Guided Vehicles (AGVs) while interacting with human workers in shared workspaces in manufacturing plants.

As a first step, we modeled worker motion as a Finite Automaton Model (FAM) with 6 intuitive states and transition functions that could be set up easily once we get the plan of the manufacturing plant. As a second step, we have added predictive modeling to the basic FAM model… Read more

I-ORCA: Implicitly Nudging Human Trajectories

In this project, we look at the inverse of the social navigation problem. Social Navigation deals with generating trajectories for mobile robots that minimally invade human trajectories.

We try to see if we can leverage the obstacle avoidance nature of humans to generate trajectories for robots that minimally nudge the humans toward desired directions… Read more

Trust Dynamics with Discounting of Performance History

In an earlier project, along with some collaborators, we found three distinct types of trust dynamics exhibited by people when interacting repeatedly with robots and automation. The figure below shows the three clusters using two distinguishing features: The root mean squared error \(E_{RMS}\) between the trust reports given by the participants and the values predicted by a performance-based trust model and the average logarithm of trust reports given by the participants.

In this project, we focus on the oscillators cluster… Read more

Course Project - EECS 545 Machine Learning - Solving Wordle using Reinforcement Learning

As a part of this course project, we used deep reinforcement learning (DeepRL) to try and solve the word game Wordle. In wordle, you have 6 chances to guess a 5-letter word. After each guess, you get feedback about the correct letters and positions until you guess the correct word.

. We implemented two versions of the model, one using a dictionary of words as an input and the other guessing one character at a time, without knowing a dictionary… Read more

Course Project - IOE 691 Bayesian Data Science - Weight Uncertainty in Neural Networks

In this project, we implemented a Bayesian Neural Network (BNN) using Python and PyTorch, with the goal of replicating the results from the influential paper, Weight Uncertainty in Neural Networks. The project involved a thorough study of the paper’s methodology, followed by hands-on development of probabilistic neural network models that explicitly capture uncertainty in weight parameters. Key skills applied include deep learning, Bayesian methods, PyTorch programming, and experimental validation through empirical replication.

Read more

Estimating Readiness of Drivers to Takeover Control of a Vehicle in Conditionally Automated Driving

I worked as a software developer for this project, helping the lead PhD student implement sensor fusion from a variety of sensors and computing some physiological metrics in real-time using the Robot operating system (ROS) in python. The iMotions software combines various sources of data (GSR sensor, eye-tracker, and heart-rate sensor) and communicates to a computer running the following ROS framework using UDP.

The computer can then compute the metrics and send it to another server running a machine learning model to predict the driver’s readiness in real-time… Read more

Course Project - IOE 536 Cognitive Ergonomics - Redesigning Venmo

In this project, we analyzed the Venmo app from the perspective of cognitive ergonomics. We found a lot of inconsistent and irrational design choices leading to bad user experience. We proposed a few redesigns that will help make the user experience better.

. The main thing we tried to remove was the ‘social media’ aspect of Venmo… Read more

Feedback Linearization Based Trajectory Planning for a Planar Two-Link Bipedal Walking Robot

The goal of this study is to perform a solution space analysis for the biped trajectory planning method using feedback linearization. As a first step, one such trajectory was generated during this phase of the study. The concept of feedback linearization was used to convert the highly complex and nonlinear problem into a relatively easy to solve linear differential equation.

At this stage, a solution to the problem that minimizes the energy input was found and in that process, all the background code that is required for the continuation of the study was set up… Read more