Ryan Diaz

Ryan Diaz

Hello! I am a first-year CS Ph.D. student at Rice University. My research interests include learning preference-aligned behaviors for robotic manipulation, and I am currently doing research with Prof. Vaibhav Unhelkar at the Human-Centered AI and Robotics Group. I'm always trying to explore new ways to help robots do cool things!

Previously, I was an undergraduate at the University of Minnesota where I worked with Prof. Karthik Desingh at the Robotics: Perception and Manipulation Lab on multisensory contact-rich robotic manipulation. I also got the chance to do research on reinforcement learning for autonomous driving with Prof. Yevgeniy Vorobeychik at the WashU CSE REU.

Outside of academics and research, I enjoy cooking, reading, and worldbuilding (as well as procrastinating way too much on actually writing). Whether it be writing papers, writing code, or writing stories, the keyboard always calls...

CV LinkedIn GitHub Google Scholar

News

Publications

hrd png

Hierarchical Reward Design from Language: Enhancing Alignment of Agent Behavior with Human Specifications

Zhiqin Qian, Ryan Diaz, Sangwon Seo, Vaibhav Unhelkar

International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS) 2026

[Paper] [Code (coming soon)]

We introduce the hierarchical reward design (HRD) problem that allows the integration of human preferences into hierarchical reinforcement learning algorithms. We instantiate a language-conditioned version of the HRD problem and solve it with an LLM-based reward generation pipeline, Language to Hierarchical Rewards (L2HR).

auginsert gif

AugInsert: Learning Robust Visual-Force Policies via Data Augmentation for Object Assembly Tasks

Ryan Diaz, Adam Imdieke, Vivek Veeriah, Karthik Desingh

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2025
Beyond Pick and Place Workshop at ICRA 2025

[Paper] [Website] [Video] [Code]

We build a multisensory imitation learning framework and evaluate it on an extensive set of task variations for a peg-in-hole task. We also explore data augmentation as a possible technique for increasing a policy's robustness to these variations.

evaluating robustness thumbnail

Evaluating Robustness of Visual Representations for Object Assembly Task Requiring Spatio-Geometrical Reasoning

Chahyon Ku, Carl Winge, Ryan Diaz, Wentao Yuan, Karthik Desingh

IEEE International Conference on Robotics and Automation (ICRA) 2024
2nd Pretraining for Robot Learning Workshop at CoRL 2023

[Paper] [Website] [Video] [Code]

We introduce a novel dual-arm object assembly task that focuses on geometric and spatial reasoning. We compare multiple pretrained vision encoders in a behavior cloning framework across a large set of grasp and object geometry variations.

Projects

robot data folder thumbnail

Filtering Robot Demonstrations with Probabilistic Data Structures

COMP 580: Probabilistic Algorithms and Data Structures (Rice University)

[Report] [Code]

We use probabilistic data structures to filter an incoming stream of robot demonstrations to maximize state-action coverage while minimizing overall dataset size.

vtmae thumbnail

Multisensory Visuotactile Pretraining for Robotic Manipulation Tasks

CSCI 5980: Deep Learning for Robot Manipulation (University of Minnesota)

[Poster] [Video]

We implement a masked pretraining objective for a vision and force-torque observation encoder and perform downstream evaluation on a series of contact-rich robotic manipulation tasks.

obstacle avoidance gif

Obstacle Detection and Avoidance in Simulated Autonomous Driving

CSE REU Project (Washington University in St. Louis)

[Report] [Poster] [Detection Code] [RL Code]

We use an object detection module and reinforcement learning to build an obstacle avoidance pipeline for an autonomous vehicle in the CARLA simulation environment.

fourier features gif

An Exploration of Fourier Features for Image and Video Representations

MATH 5466: Mathematics of Machine Learning and Data Analysis II (University of Minnesota)

[Report] [Code]

We investigate the effectiveness of Fourier Features in MLPs for coordinate-based representations of images and videos. We also explore their theoretical motivations via the Neural Tangent Kernel (NTK).

political bias in news thumbnail

Temporal Political Bias Analysis in Media Using Language Models

CSCI 5541: Natural Language Processing (University of Minnesota)

[Report] [Poster] [Code]

We use a pretrained language model to analyze the political bias of articles from various news sources over an extended period of time.

nerf sam seg gif

3D Semantic Segmentation of a Scene Using NeRFs and SAM

CSCI 5561: Computer Vision (University of Minnesota)

[Report]

We perform semantic segmentation of a 3D scene by applying a Segment Anything Model (SAM) to a NeRF-rendered scene from a series of 2D images.

slot diffusion policy gif

Object-Centric Representation for Imitation Learning in Robotics

CSCI 5527: Deep Learning: Models, Computation, and Applications (University of Minnesota)

[Report] [Code]

We explore the use of slot-based representations (via slot attention) in the diffusion policy framework for robotic manipulation tasks.

blender obj generation thumbnail

Programmatic Object Generation with Blender

Undergraduate Research Scholarship Project (University of Minnesota)

[Report] [Poster] [Video]

We utilize the Blender Python API to generate a large dataset of objects with wide varieties of geometries for use in simulating a peg-in-hole task.

freecell ai thumbnail

Using Search Algorithms to Solve Free-Cell Solitaire Games

CSCI 4511W: Introduction to Artificial Intelligence (University of Minnesota)

[Report] [Code]

We compare both uninformed and informed search algorithms (using a wide variety of heuristics) for solving a scaled-down version of the free-cell solitaire game.

wrsct gif

Weather Radar Snow and Storm Cell Detection

Personal Project

[Code]

We use classical computer vision techniques to segment snow and storm cells from weather radar images.