RoboArena is a distributed (7+ partner institutions) real-world generalist robot policy benchmark. It aims to provide comprehensive evaluation of user submitted robot policies in diverse real-world scenarios. It is built on the idea that we can scale real-world evaluations in both number and diversity if we let robot evaluators pick any scene and task for each evaluation, and run evals A/B style. Check out the website for more details!

Evaluating robot policies is a laborous, time-consuming process, and so far needs to be done by humans. It turns out if we train specialized reset policies and a success classifier, we can automate away all human involvement. Evaluations are still done in the real world, so they are trustworthy and accurate, but now they're also fully autonomous and scalable.

Robot learning requires data, but does the data need to be human collected? We find that we can boostrap a self-improvement process with a pre-trained policy and VLMs, collecting autonomous data at scale that can be used to improve the policy completely on its own.

What prompts elicit language models to best answer knowledge intensive questions? We find that a mix of in-context examples that the model knows how to answer and doesn't know yields optimal QA performance.

SuSIE leverages the internet pretraining of image generation models like InstructPix2Pix to achieve zero-shot robot manipulation on unseen objects, distractors, and scenes.

We devise Optim-FKD, a new paradigm for accurate, high speed control of a robot using a learned forward kinodynamic model and non-linear least squares optimization.

In this work we learn a visual inverse kinodynamic model conditioned on image patches of the terrain ahead to better enable high-speed navigation on multiple different terrains.

Optimal sampling-based motion planning algorithms when applied to kinodynamic planning make a trade off between computational efficiency and solution quality. With S3F we demonstrate that both are attainable by proposing a new way to learn the steering function required by these sampling-based planners.

Built a functional OS from scratch with preemptive scheduling, virtual memory, file IO, and system calls.

Implemented the 2PC protocol for execution of a distributed atomic transaction. Uses memory safety and concurrency features of the Rust language.

Simulated 2D and 3D fluid dynamics using the Navier-Stokes equations. Computation was accelerated with GPU code written in CUDA.

DyslexiAR helps dyslexic individuals read and write by recognizing words via OCR and displaying augmented reality models representing the words.

Wrote an interpreter and compiler for a programming language with data types, dynamic memory allocation, functions, control structures, and file IO. Compiler is self-hosting (the compiler is written in the language it compiles).

Implemented an efficient neural network training library from scratch. Library implements many deep learning primitives such as various activation functions, L1 and L2 regularization, and dropout.

Developed software that crawls web pages and efficiently indexes them using a compressed trie. Built a search engine using relevance feedback and the PageRank algorithm.

Implemented a five stage processor in Verilog based off a custom ISA. Can handle data, resource, and control hazards. Implemented a branch predictor with a direct-mapped cache.

Fitted Raspberry Pi with ultrasound sensor to take ultrasound scans of the body part in question. The Pi is programmed to process scan image with a CNN and determine the likelihood of the presence of five different deleterious growth types.

Analyzes biases of the publisher, recency of the articles, and performs fact checking with other news sources to evaluate credibility of news article.