Sharmistha S. Gupta

Multi-Agent Robot Learning and Control

Human-Robot Collaboration using Imitation Learning

Enabling robots to collaborate with humans requires both predicting the human's behavior, and determining appropriate robot responses. Imitation learning, or learning from demonstrations of human-human collaboration is a sample efficient way to achieve the same. Prior work on human-robot collaboration using imitation learning is mostly centered on the underlying assumption that human behaviours in a particular situation can be narrowed down to a specific expert policy. On the other hand, research that deals with varying human behaviour does not talk about scenarios where humans and robots collaborate without direct communication. We are working on a novel framework to train an agent that must learn to make intelligent observations and take sequential decisions in order to collaborate effectively with a human and work towards the same goal, while taking into account the latent factors present in expert demonstrations.
Work Referred:

InfoGAIL: Interpretable Imitation Learning from Visual Demonstrations [paper]

Multi-Agent Generative Adversarial Imitation Learning [paper]
Advisors: Dr. Katia Sycara, Dr. Dana Hughes

Multi-Agent Swarm Herding Using Reinforcement Learning

Swarms are decentralized systems wherein multiple agents follow common control laws and behave as a collective entity. In nature, swarm behaviour can be observed in various animals, insects, fish etc. for foraging, or navigation purposes. We consider Strombom's shepherding model which was proposed in 2014 to model collecting and herding actions of a shepherd (predator) driving sheep (agents) towards a destination. We employ reinforcement learning techniques to achieve better results as compared to the proposed heuristic approach by Strombom et al.
Work Referred: Solving the shepherding problem: heuristics for herding autonomous, interacting agents [paper]
Advisor: Dr. P. B. Sujit

Leader Placement in Robotic Swarm Control

Human control of a robotic swarm entails selecting a few influential leaders who can steer the collective efficiently and robustly. We are conducting experiments to vary number of leaders and leader positions in simulated robotic swarms of different sizes, and assess their effect on steering effectiveness and energy expenditure. We are analyzing the effect of placing leaders in the front, middle, and periphery, on the time to converge and lateral acceleration of a swarm of robotic agents as it performs a single turn to reach the desired goal direction. The basic results are published in the paper mentioned below. We are conducting follow up experiments for the same before submitting the work to a journal shortly.

Work Referred: Effect of Leader Placement on Robotic Swarm Control [paper]
Advisor: Dr. P. B. Sujit, Dr. Sachit Butail

Robotic Systems

Virtual Reality Based Control Of Robotic Swarms

Robotic swarms do not have intelligence for high level tasks. We need human input for effective mission execution and also for applications where physical presence of a human isn’t possible. Virtual reality provides an interactive environment between humans and robots, in a way that is very effective and close to real world scenarios. We used an HTC Vive headset and a Unity environment to control a shepherd model of a swarm, led by a SkyViper v2450GPS quadrotor. Due to infrastructure issues, the remainder of the swarm was simulated virtually. We were able to successfully control the leader virtually and observe the motion of the swarm accordingly. We will be conducting a few more experiments before submitting the work shortly.
Work Referred: Human-Robot Interaction: A Survey [paper]
Advisor: Dr. P. B. Sujit

Fisheye Vision in Self-Driving Cars

The aim of this project was to employ imitation learning and use a fisheye camera as opposed to a set of three frontal cameras (left, right and center)for improving vision in self driving cars. Data was collected in the form of humans controlling the ground rover and the human steering angle was the only training signal. We built the entire system hardware on our own using an Nvidia Jetson TX2 module as its brain. The camera used was Point Grey Chameleon 3. We had preliminary results from a previous dataset. However, due to infrastructure and resource issues, we had to temporarily discontinue the project.
Work Referred: End to End Learning for Self-Driving Cars [paper]
Advisor: Dr. Chetan Arora

Human-Technology Interaction

Perception of Fake News on Social Media

We attempted to understand the perspectives of the seniors (aged 65 years and above) on misinformation in the Indian context. Interviews with 33 seniors who use social media regularly revealed three themes. The seniors viewed and rationalized sharing news irrespective of its veracity as a process of building sociality. Sharing information was also based on the logic of superimposing information with an epistemic ascription to the networks from where they received it. Finally, a kind of normative dualism becomes apparent from an acknowledgment of the role they may play in the spread of misinformation as agents on the one hand and a resounding need to stop it on the other due to its potential social ramifications.
Publication: Building Sociality through Sharing: Seniors' Perspectives on Misinformation
Advisor: Dr. Shriram Venkatraman

Digital Payment Platforms for Financial Inclusion

Financial transactions are intimately bound with social interactions and woven into everyday economic lives. We focused on PayTM, a digital wallet, and a specific set of users of PayTM, street vendors in urban India. Through an ethnographic investigation we unpacked two questions: 1. Can digital forms of money create financial inclusion by opening up access to the marginalized 2. Can digital platforms amplify socio-economic capacities of low literate users enhancing financial literacy? We argue that digital and financial literacy are an immersive component of digital wallet use acquiring ‘everyday life’ in specific socio-economic ecosystems. Our study captures daily practices of digital money staking a claim in advancing the understanding of financial inclusion as a lived process accumulating habits, practices and stakes to expand socio-economic capabilities.
Publication: Digital Wallets ‘Turning a Corner’ for Financial Inclusion: A Study of Everyday PayTM Practices in India
Advisor: Dr. Nimmi Rangaswamy