Sateesh Kumar

I am a second year Computer Science PhD student at The University of Texas at Austin. My research focuses on the intersection of Robotics and Computer Vision. I hold a Master's degree from The University of California, San Diego where I was advised by Prof. Xiaolong Wang and a Bachelor's degree from FAST NUCES Karachi.

In addition to my academic pursuits, I have gained industry experience as a Research Intern at Amazon Fauna Robotics (Bellevue), mentored by Cheng-Hao Kuo and Sadegh Rabiee, as a Generative AI Researcher at TikTok, and as a Computer Vision Research Engineer at Retrocausal.

Email  /  CV  /  Google Scholar  /  Twitter

• Interned at Amazon Fauna Robotics, Bellevue, mentored by Cheng-Hao Kuo and Sadegh Rabiee (Summer 2026).
• Awarded the Amazon AI PhD Fellowship (2025-2027) (August 2025).
• COLLAGE accepted to CoRL 2025 (Aug 2025).
• Started PhD (CS) at the University of Texas at Austin (Aug 2024).
• Won the Datacomp challenge at ICCV 2023 (Oct 2023).
• Completed MS in Computer Science at UCSD and joined TikTok/ByteDance Seed team as a Research Engineer (June 2023).
• GraphIRL accepted for oral presentation (top 6%) to CoRL 2022 (Aug 2022).
• Interning at TikTok/ByteDance (June 2022).
• TOT accepted to CVPR 2022 (Feb 2022).
• LAV accepted to CVPR 2021 (Feb 2021).

*: Equal Contribution. †: Equal Advising.

MoMaSplat is a demonstration augmentation framework for mobile manipulation that synthesizes full manipulation trajectories, including the moving object, from unseen robot base poses. It reconstructs the scene and manipulated object using 3D Gaussian Splatting and feed-forward 3D foundation models, and retargets actions via kinematics, enabling policies to succeed from base poses beyond 25 cm where all baselines fail.

COMPASS selects base poses for deploying VLA policies on mobile manipulators by using the policy's own predictive uncertainty as a measure of competence. It renders candidate observations from a 3D Gaussian Splatting reconstruction of the scene and picks the pose where an ensemble of VLA checkpoints agrees most, outperforming the strongest baseline by 35 percentage points on a real-world humanoid without needing training data or execution labels.

COLLAGE is a few-shot imitation learning method that adaptively fuses demonstrations from multiple similarity modalities. It estimates each modality’s usefulness by training a policy on its retrieved data and measuring how probable the target actions are under that policy.

MimicDroid enables humanoid robots to learn manipulation tasks from freely recorded human play videos using in-context learning, achieving nearly twofold higher real-world success rates compared to existing approaches without requiring expensive teleoperated demonstrations.

STAR (SpatioTemporal Active Retrieval) enables service robots to locate objects in dynamic environments by combining non-parametric long-term memory with a working memory, using vision-language models for unified spatial and temporal decision-making.

This paper introduces a method that uses finetuned multimodal language models to filter image–text pairs more accurately than traditional metrics like CLIPScore. It defines multiple specialized quality metrics and builds instruction tuning data guided by stronger models such as GPT-4 to train the models to score data effectively. The result is a reliable and efficient filter that better evaluates image–text alignment.

We introduce a three-stage filtering strategy for enhancing model performance. It focuses on single-modality filtering, cross-modality filtering, and data distribution alignment. The proposed approach significantly surpasses previous methods on the DataComp benchmark.

GraphIRL is a self-supervised method for learning a task reward solely from videos. We build an object-centric graph abstraction from video demonstrations and then learn an embedding space that captures task progression in a self-supervised manner by exploiting the temporal cue in the videos.

We propose temporal optimal transport for jointly learning representations and performing online clustering in an unsupervised manner. The approach learns prototype vectors via backpropogation. The prototype vectors are initialized at random and act as cluster centroids.

We propose alignment as pre-text task for self-supervised video representation learning. The proposed approach leverages differentiable dynamic time warping for learning global alignment across pairs of videos.

We collect a video dataset of road-based anomalies. We propose an object-object interaction reasoning approach for detecting anomalies without additional supervision. We experiment with reconstruction based and one-class classification based approaches.

Website layout is from Jon Barron