Sandeep Routray

I'm a Master's student in Machine Learning at Carnegie Mellon University, where I work with Prof. Deepak Pathak on scaling robot learning using internet-scale videos without action labels.

Previously, I was at Samsung Research, Seoul, where I developed Map View, a framework to extract vectorized home layouts from in-the-wild floorplan images - our solution was showcased at CES 2024. I also spent time as a research fellow at the Vector Institute for AI, working with Prof. Sanja Fidler on object-centric self-supervised learning by leveraging inter-image relationships.

I graduated from Indian Institute of Technology Kanpur with a major in Electrical Engineering and a minor in Machine Learning. During my undergraduate studies, I worked on projects in natural language processing, optimization algorithms, stochastic modeling, and wireless networks.

Email  /  Resume  /  Scholar  /  Twitter  /  Github

Affiliations

Research

I am broadly interested in building general-purpose embodied agents that can operate in unstructured real-world environments. I work on the following core themes:

• Leveraging Internet-scale video and language data for scalable, label-efficient robot learning.
• Using pretrained multimodal models to enable knowledge transfer across tasks and embodiments.
• Leveraging video generation models to learn predictive world models that can be adapted for policy learning and planning.
• Learning actionable representations from complex scenes that integrate multiple modalities such as images, videos, depth maps, and 3D data.

Publications

Developed a system for SemEval 2020 Task 4 on commonsense validation and explanation using pretrained language models. Our approach leverages parallel model instances to jointly reason over sentence pairs and identify contradiction cues. It achieved 94.8% accuracy on subtask A (classification) and 89% on subtask B (explanation selection), ranking among the top-performing submissions.

Analyzed spectral efficiency (SE) in a K-tier heterogeneous cellular network modeled using Poisson point processes (PPP) and Poisson cluster processes (PCP). We derived Laplace transforms of received power and interference, enabling tractable SE expressions and insights into the impact of user clustering on network performance.

Projects

Proposed a slot-based inter-image contrastive learning framework for dense self-supervised representation learning. The method extends dense clustering objectives with prototype-driven context vectors to mine semantically meaningful inter-image positives and negatives. Demonstrated improved unsupervised segmentation performance on COCO and PASCAL, highlighting the value of inter-image supervision for part-aware, generalizable representations.

Borrowed from Jon Barron website's source code. Do not scrape the HTML from this page itself, as it includes analytics tags that you do not want on your own website — use the github code instead. Also, consider using Leonid Keselman's Jekyll fork of this page.