projects | Oliver Wang

Current research

Robot memory on a real quadruped

Sep 2026 – now · NESL · Unitree Go2

What breaks when a routine-aware memory leaves simulation and runs on a real robot?

I deployed our memory system on a Unitree Go2, built memories from 3,000+ robot images, and compared five memory designs. Real-world failures that never showed up in simulation included inconsistent scene descriptions, localization drift that corrupted stored observations, and memories that could not represent something being absent. I also built an agent that diagnoses robot failures from tool-call and sensor logs.

Home robot memory when routines change

Jun 2026 – now · NESL · CoRL 2026 workshops

When people start keeping things in new places, how should a home robot update what it remembers?

I built a ten-household object-search benchmark where routines change without warning. LLM-written memory summaries often fail to reuse what they learned when an earlier routine comes back. My memory-update method keeps hypotheses about the underlying routines and uses a retrieval tool to check them against raw observations, which improves predictions during these shifts, especially recurring ones.

Confidence after a routine shift

Jun 2026 – now · NESL · CoRL 2026 workshops

After someone's routine shifts, does a predictor's confidence still tell you when it is wrong?

I compared LLM self-reported confidence with ensemble uncertainty after routine changes. LLM confidence is poorly calibrated, but it keeps a ~0.3 correlation with correctness while the baselines fall to about zero.

Finding events in cameras, sensors, and text

Jan 2026 – now · with Sandia National Labs

Can an agent keep revising what it believes is happening as evidence about an unfolding event arrives?

I lead the research direction of this UCLA–Sandia collaboration. I proposed a knowledge-graph-based agent that maintains and revises hypotheses about urban events, like wildfires and traffic accidents, as evidence comes in from cameras, sensors, and social media. I'm now building a version that refines its own memory from detection outcomes, together with a master's student I recruited to the project.

Alzheimer's detection from speech

Sep 2026 – now · SPAPL, UCLA (Prof. Abeer Alwan)

Can the confidence and disagreement of speech recognizers help detect Alzheimer's disease?

I'm using calibrated confidence and disagreement across multiple speech recognizers as features for detecting Alzheimer's disease from picture-description recordings, and developing an agentic method that uses clinical knowledge to choose between feature types when there are very few training samples.

Human feedback for a time-series benchmark

Dec 2025 – now · NESL

How do you check that a synthetic dataset teaches a time-series model the right things?

I built a review pipeline where 10 annotators went through 1,000 questions, and worked with the lead researcher to fix ambiguous tasks, questions, and labels in a synthetic dataset of 10,000+ samples for training a general-purpose time-series foundation model.

Spectral predictability

Jan 2024 – Dec 2025 · NESL · AAAI 2026 AI4TS workshop, oral

Before running a large forecasting model, can you tell whether it will beat a small one?

A spectral predictability score takes seconds to compute. Across 51 models and 28 datasets, foundation models had 20–60% lower error than lightweight forecasters on data the score rated as predictable.

arXiv code

Earlier research

Robots reading recipes

Summer 2023 · Cross Labs, Kyoto

Can an off-the-shelf language model turn a written recipe into code a robot can run?

Working with cocktail recipes and no fine-tuning, I combined line-by-line prompting, ensemble voting, and verb parsing to reach 0.75 accuracy. I also wrote feasibility reports for industry clients.

paper

Insertable glucose sensor

2023 – 2024 · Ozcan Research Group, UCLA

Can a cheap, compact imager read glucose levels from a sensor under the skin?

I calibrated the phosphorescence lifetime imager and built segmentation and neural network models to cut down variability between readings.

arXiv

Layer-scaled one-shot pruning

2022 · high school

How much of a neural network can you remove in one pass?

My first research paper, on a one-shot pruning method for making neural networks smaller.

paper

Things I built

FPGA audio visualizer

2024 – 2025 · UCLA IEEE, ML lead

Can an instrument classifier run in real time on a small FPGA?

I led the team that built a real-time audio visualizer, and trained a 12-class instrument classifier (above 70% accuracy) that we deployed on a Xilinx FPGA.

Micromouse

2022 – 2023 · UCLA IEEE · 5th place, All America Micromouse

Can a small robot we designed ourselves map and solve a maze on its own?

We designed the PCB and IR sensors and wrote the flood-fill solver and firmware.

Arduino instrument

Side project

Can a musical keyboard teach how binary encoding works?

A keyboard-style Arduino instrument built to teach binary encoding schemes.