Akash Karthikeyan

MASc Candidate, Electrical & Computer Engineering, University of Waterloo

Hello There!

I'm a 2nd-year master's student affiliated with the CL2 Lab, working on generative models (diffusion) for generalizable path planning. I am currently advised by Prof. Yash Pant.

Previously, I was a MITACS GRI at the University of Toronto, where I worked on novel view generation and animation of human models (NeRFs + diffusion) under the guidance of Prof. Igor Gilitschenski. I completed my Bachelor’s in Mechanical Engineering at TCE, focusing on fast, depth-aware grasping for transparent objects under the guidance of Prof. Saravana Perumaal.

Akash Karthikeyan's profile photo

Email  |  CV  |  Kaggle  |  GitHub  |  Google Scholar  |  LinkedIn

I'm looking for PhD positions starting Fall 2025

I'm interested in leveraging generative models to tackle core challenges in generalization and robustness (robotics + 3DV).

  • Can we use differentiable logic constraints to enhance realism and robustness, and how does entropy-based constraints influence these methods?
  • Can we exploit 3D biases or learn disentangled representations to improve generalization in these models, enabling the development of a world model?
  • Is it possible to achieve the above goals in a sample-efficient manner?

Highlights

[Mar. 2024] Presented our work Adaptformer at AIA, Stanford University, CA, USA (AIA 2024)
[Feb. 2024] Presented our work AvatarOne at the Vector Research Institute (Remarkable - Symposium)
[Jan. 2024] Presented our work AvatarOne at WACV 2024
[Feb. 2023] Best Outgoing Student in BE Mechanical Engineering (Department Gold Medal)
[Jul. 2023] Runners-up| Certifiable 3D Registration, (NCVPRIPG 2023)
[Aug. 2022] Achieved Kaggle Competitions Expert status, ranked 232 globally!
[Aug. 2022] Winner | Tamil Nadu Student Innovator Award

Research

AdaptDiffuser: Energy Diffusion for Generalizable Planning

Akash Karthikeyan, Yash Pant

Under Review, 2024

An energy diffusion based model for adaptable and generalizable planning.

Adaptformer: Sequence Models as Adaptive Iterative Planners

Akash Karthikeyan, Yash Pant

Under Review, 2024

Project Page / Video

Transformers as adaptive planners through iterative energy minimization.

AvatarOne: Monocular 3D Human Animation

Akash Karthikeyan, Robert Ren, Yash Kant, Igor Gilitschenski

WACV, 2024

Project Page / Paper / Code / Video

Fast novel view generation and animation of human avatars from monocular views only.

CAMM: Building Category-agnostic and Animatable 3D Models from Monocular Videos

Tianshu Kuai, Akash Karthikeyan, Yash Kant, Ashkan Mirzaei, Igor Gilitschenski

CVPRW, 2023

Project Page / Paper / arXiv / Code / Data

Kinematic aware, animatable, and class-agnostic 3D deformable objects from casual monocular videos without the use of any priors or camera poses - A Template-Free Approach

Competitions and Projects

*Below are some of the projects I had fun working on, click the title for details.

Lyft L5 toolkit - Motion Forecasting

Kaggle - Build motion prediction models for self-driving vehicles | 2021

Motion prediction with rasterized images using CNN and LSTM encoder-decoder models. Also exploring Kalman filters and other optimization techniques to enhance non-learning-based approaches. Currently working on Vectornet for a faster, raster-free pipeline.

Segment and Label Helmets in Video Footage

Kaggle - NFL, 2021

Detection of helmets, player classification, tracking using Kalman and DeepSORT for accurate mapping. Implemented predictive tracking and analysis of game trends using past data.

Automated Annotation and Classification of Catheters in Chest X-Rays

Akash Karthikeyan, Saravana Perumaal Subramanian

ICCCSP, 2022

paper

Annotate and classify catheter positions with a robust approach that accurately identifies endpoints, despite their footprint being less than 1% of the overall image.

Google Decimeter Challenge

Kaggle | Google, 2022

Results and Architecture

• Enhanced GNSS positioning and navigation accuracy on smartphones.
• Processed and cleaned GNSS logs to compute precise location, achieving a 94th position out of 810.
• Implemented Kalman smoothing and velocity-based models for improved GNSS data accuracy.

Action Recognition and Tracking of People in Real-time

slides

Winning Solution | NVIDIA - Academia Connect Hackathon, 2021

End-to-end detection, pose estimation, and tracking with ReID for crowded environments under varied lighting conditions.

Bag-Valve-Mask Ventilator

Slides

• Designed a slider-crank actuation mechanism to automate compression in BVM ventilators.
• Conducted pressure trajectory analysis using COMSOL multiphysics to achieve required PIP and PEEP values at the outlet.

Zero Energy Air-Cooler

Winner | TNSI - 2022 (certificate)

• Fabricated and characterized a device utilizing MOF as advanced moisture sorbents for energy-efficient high-temperature cooling.
• Conducted XRD and Thermogravimetric Analysis to assess MOF structure and water uptake characteristics.

More Projects

  • G2Net Gravitational Wave Detection - European Gravitational Observatory: Find gravitational wave signals from binary black hole collisions Used FFT and 1D CNN to detect the gravitational waves.
    Placed 33/1219 teams
  • March Machine Learning Mania 2022 - Men’s: Predict the 2022 College Men's Basketball Tournament (kaggle competition), also here is the visulalization tool for bracket builing NCAA_BracketBuilder
    Placed 14/930 participants
  • Crypto- Market prediction based on anonymised features (Ubiquant Market Prediction) I Used Tabnet based model.
    Placed 99/2893 participants

TEAM YUKTA RACING - Autobots Assemble!
We Build, We Race.....! Defending Champions of GKDC 2020


I borrowed this template from Jon Barron's website.