Anant Rai

Research Engineer
1x Technologies

Email  /  GitHub  /  LinkedIn  /  CV

In this project, we addressed a key barrier in home robotics, particularly in imitation learning: the lack of affordable and user-friendly tools for collecting robot demonstrations. To solve this, we developed an innovative, cost-effective tool named the Stick using which we compile the 'Homes of New York' (HoNY) dataset, encompassing 13 hours of interactions from 22 NYC homes, complete with comprehensive video and action data. We further proposed the Home Pretrained Representation (HPR), a ResNet-34 model pre-trained on the HoNY dataset using self-supervised learning, to initialize robot policies in new environments. We show that this initialization can beat baseline models in unseen home environments.

This work explores the potential for consolidating robotic learning by developing "generalist" X-robot policies adaptable across various robots, tasks, and environments. Traditionally, robotic learning involves training separate models for each application, robot, and environment. We are able to show that training on largescale diverse robotic dataset can be used to learn general skill like better spatial understanding in both the absolute and relative sense, generalization to unseen objects and overall a much more robust model that can beat baselines trained on robot-specific/task-specific data only

In this work, we present Fast Imitation of Skills from Humans (FISH), a new imitation learning approach that can learn robust visual skills with less than a minute of human demonstrations. Given a weak base-policy trained by offline imitation of demonstrations, FISH computes rewards that correspond to the “match” between the robot’s behavior and the demonstrations. Along with beating the state-of-art, FISH is versatile, which allows it to be used across robot morphologies (e.g. xArm, Allegro, Stretch) and camera configurations

Built a module using SSD (45 mAP) and SORT to detect and track agents in novel video dataset of Indian Traffic. Developed special LSTM based network architecture to learn the agent’s behaviours and interactions for trajectory prediction in dense heterogeneous traffic. Devised Weighted-Elliptical-Model for risk modelling and combined it with trajectory prediction to get novel predictive-risk-analysis (20% improvement over baseline)

Unlabelled videos of humans are a large and almost untapped source of data for robots to learn several routine tasks. The paper RLV proposes a framework that can leverage such videos to enhance learning of challeneging vision-based tasks. The drawback to their model is that it can be extremely unstable and lead to a lot of zero reward rollouts. In our work, we make the framework more robust by adding motion understanding module to stabalize cross domain adaption using discriminator. In our experiments on the task of Visual Puahing, we see our framework doing better than the RLV (15% improvement in performance)

Devised optimal environment representation using trees and graphs for intelligent cost calculation and re-routing decisions in the environments of NeurIPS2021: Flatland challenge . Using Conflict Based Search (CBD) and Deep-Q learning achieved decent performance on the environment (3rd position in round 1 and 6th position on round 2)

Used Data Mining and an array of Deep Learning techniques like Collaborative filtering, NN classifier and Sentiment Analysis to analyse and capture food habits of Indian customers. Collected and organised data from food giants like Swiggy and Zomato on a flask-server, for analysis and predictions based on location, and presented results on a web-app created using Node Js.