Viswesh Nagaswamy Rajesh

Jun '24

May '24

Dec '23

May '23

Mar '23

Jan '23

Dec '22

Dec '22

May '22

Apr '22

Mar '22

May '21

Worked under the supervision of Prof. Igor Gilitschenski at the Toronto Intelligent Systems Lab, UofT. Here, I worked on developing a two stage pipeline for adaptive online dynamics state estimation in unforseen environments.

Worked under the supervision of Prof. Masayoshi Tomizuka. Proposed a novel metric called the Influence Index to quantify coordination levels in multi-agent reinforcment learning.

Worked under the supervision of Prof. Shishir Kolathaya StochLab, IISc on implementing algorithms for quadruped locomotion. I had the opportunity to work with real-world robots and benchmarked algorithms on the Stochlite quadruped.

Worked with the Computer Vision team under the supervision of Mr. Pradeep Janakiraman on building a pipeline for real-time tracking of basketballs using YOLOv5.

Worked under the supervision of Dr. Debashish Chakravarty at AGV, IIT Kharagpur on computer vision and multi-agent systems. As a part of the group, I collaborated on the University Rover Challenge, Indy Autonomous Challenge, and the Machine Learning Reproducibility Challenge with the group. I also had the opportunity to lead the deep learning and vision team for two years.

Under Review at Confenrece on Robot Learning 2024 (CoRL)

More details to be shared soon

Proposed Entity Augmentation, a novel approach that eliminates the need for private set intersection (PSI) and entity alignment in Vertical Federated Learning for categorical tasks

Reproduced the results of the seminal YNet architecture of social trajectory prediction. Proposed a novel transfer learning experiment that achieved state-of-the-art performance on the SDD dataset.

Member of the Gold Winning Team, IIT Kharagpur

Adobe Behavior Simulation Challenge: The task was two fold: (i) Given the contents of a tweet and any media files along with username, predict the number of likes (ii) Given the number of likes, as well as the username and any media files, predict the text contents of the tweet.

Our approach involved finetuning LLama, NeXT-GPT and LLaVA 1.5 on the dataset. We leveraged LLaVA for video-captioning followed by keyword extraction to engineer a custom prompt. We optimized the pipeline further using a bandit-informed routing algorithm to select the best LLM during inference. [code]

Member of the AGV, IIT Kharagpur team

The task was to reproduce research papers from top AI conferences and extend the ideas for state-of-the-art performance. We successfully reproduced the results of "From Goals, Waypoints & Paths To Long Term Human Trajectory Forecasting" (ICCV '21). Further, we proposed a novel transfer learning experiment that achieved SOTA performance on the Stanford Drone Dataset (SDD) and the Intersection Drone Dataset (InD)

The work was accepted at ReScience C Journal and we presented a poster at NeurIPS 2022 [code]

Member of the IIT Kharagpur - IUPUI, Indiana - USB Colombia collaborative team.

Designed tightly/loosely coupled high-speed localisation in for racecar localisation in pre-mapped LiDAR circuit. The localisation was using 3 static-state LiDARs.

Wrote ROS packages to construct the map from LiDAR scans, extract the local map and to implement Iterative Closest Point algorithm for localization in mapped environment. [code]

Implemented a two-stage pipeline for online dynamics estimation in unforseen environments. Proposed a novel encoder-decoder architecture for the first stage and performed real-time experiments on a 1/10th scale car. Work is under review at Conference on Robot Learning 2024 (CoRL)

Explored novel methods to quantify the coordination between two agents on the Meltingpot environments.

Proposed Influence Index, a scalar metric that measures the interaction level between agents in varying environments. Implemented Self-Play, Population-Play and Fictitious Co-play methods and obtained results.

Worked on the ROS control framework of the Stochlite Robot and benchmarked the Soft Actor-Critic and Advantage Actor-Critic algorithms on the same.

Explored gradient free methods such as augmented random search for end foot trajectory spline generation.

Implemented the DQN, DDQN, PPO and TD3 algorithms to solve the LunarLanderv2 and BipedalWalker-v3 environments within OpenAI Gym. Explored gradient clipping, reward normalization and advantage estimation (GAE) to achieve rewards of over 200.