Dhruv Sheth

I’m an undergraduate student at the California Institute of Technology (Caltech) studying Computer Science and Robotics. I’m currently working with Prof. Georgia Gkioxari in collaboration with AMBER Lab on improving 3D Vision Priors in Robotics. At Caltech, I’m fortunate to be advised by Prof. Soon Jo Chung and Prof. Joel Burdick.

I previously spent time as a ML intern at Persona AI, where I worked on robotics infra and explored scaling effects of data and model size on VLA generalization, at Edge Impulse (acqd. by Qualcomm), where I focused on ML model quantization for embedded systems, and at Luxonis, where I worked on the SpatialAI library for on-device computer vision.

My research interests include long-horizon reasoning in robotics and probing generalization in Vision Language Action (VLA) models.

news

Sep 22, 2025	Our paper “SpatialTraceGen: High-Fidelity Traces for Efficient VLM Spatial Reasoning Distillation” was accepted to NeurIPS 2025, Efficient Reasoning Workshop.
Jun 10, 2025	Joined Persona AI as an ML intern this summer to work on VLAs and scale robotics infra.
May 20, 2025	Awarded Second Place at Robotics Grasping and Manipulation Competition (RGMC), In-Hand manipulation track at ICRA 2025.
Jun 16, 2024	Awarded Dr. Jane Chen Summer Undergraduate Research Fellowship (SURF) to purse research at Dr. Joel Burdick’s Robotics Lab on the DARPA Learning Introspective Control (LINC) research program
Sep 18, 2023	Moved to Pasadena to start my undergraduate studies at Caltech!

selected publications

SpatialTraceGen: High-Fidelity Traces for Efficient VLM Spatial Reasoning Distillation

Gio Huh^*, Dhruv Sheth^*, Rayhan Zirvi^*, and Frank Xiao^*

Efficient Reasoning Workshop, NeurIPS, 2025

PDF Code Dataset
ElastiCL: Elastic Quantization for Communication Efficient Collaborative Learning in IoT

Bharath Sudharsan, Dhruv Sheth, Shailesh Arya, Federica Rollo, Piyush Yadav, Pankesh Patel, John G. Breslin, and Muhammad Intizar Ali

In Proceedings of the 19th ACM Conference on Embedded Networked Sensor Systems, 2021

Abs DOI PDF

Transmitting updates of high-dimensional models between client IoT devices and the central aggregating server has always been a bottleneck in collaborative learning - especially in uncertain real-world IoT networks where congestion, latency, bandwidth issues are common. In this scenario, gradient quantization is an effective way to reduce bits count when transmitting each model update, but with a trade-off of having an elevated error floor due to higher variance of the stochastic gradients. In this paper, we propose ElastiCL, an elastic quantization strategy that achieves transmission efficiency plus a low error floor by dynamically altering the number of quantization levels during training on distributed IoT devices. Experiments on training ResNet-18, Vanilla CNN shows that ElastiCL can converge in much fewer transmitted bits than fixed quantization level, with little or no compromise on training and test accuracy.