Samrat Sahoo
I am currently a Master’s student at Stanford University, where I’ve been fortunate enough to work on robot learning with Jeannette Bohg and Dorsa Sadigh. I have also had the opportunity to collaborate with Tom Silver at Princeton PRPL.
My research is interested in answering (1) how can we achieve large-scale robot data collection with limited human supervision and (2) what makes some robot data more useful than others, and how can we generate and use data with those properties for downstream policy learning? Before Stanford, I studied computer science at Georgia Tech, where I specialized in artificial intelligence and low-level systems.
I am applying to PhD programs in Robotics for Fall 2027.
Research
* equal contribution
-
Under review at ICRA 2027
TANDEM: Task and Motion Planning with As-Needed Demonstrations for Efficient Vision-Language-Action Model Fine-tuning
TANDEM (Task and Motion Planning with As-Needed Demonstrations) is a system that combines task and motion planning with selective human teleoperation to collect demonstrations for tasks beyond a planner’s capabilities.
-
Under review at ICRA 2027
DATAFARM: Distribution-Aligned Task and Motion Planning for Fine-Tuning Vision-Language-Action Models
DATAFARM uses the pretraining distribution to guide task and motion planning trajectory generation, aligning joint configurations, motion style and timing.
-
Under review at ICRA 2027
One Demonstration, Many Objects: Generalizing Manipulation via Local Contact Geometry
DemoMimic learns dexterous manipulation from a single human demonstration, with one policy per task that transfers across objects of different shape, scale, mass, and friction.
-
IEEE International Conference on Blockchain · 2024
Scatter Protocol: An Incentivized and Trustless Protocol for Decentralized Federated Learning
A blockchain protocol for decentralized federated learning that uses staking, incentives, and penalties to deter malicious participants.
