Category: Paper
-
New preprint: latent-entropy anomaly detection
In collaboration with ECE colleagues, we have put out a new variant of our unsupervised anomaly detection pipeline. This one uses a latent entropy loss to scramble the latent space, making anomalies harder to reconstruct (and hence easier to detect). No supervision (including a normal-only dataset) needed! Citation:
-
New preprint: verifiable deterministic world models
Our exploration of world models for system assurance resulted in a semi-predictable but currently unfashionable choice: removing randomness and uncertainty from the latent space made world models more verifiable (although a tiny bit less picture-perfect). More surprisingly, this step made the behaviors produced by them more relevant to the real world. As a result, we…
-
New preprint: statistical-symbolic verification of perception
Our collaboration with RPI has yielded an extended and improved version of our NeuS’25 paper: combining conformal prediction for neural perception with reachability analysis for the dynamics and control. This problem required constructing a discrete abstraction of the perception neural net, which we did with a genetic algorithm. Citation:
-
New preprint: a survey of CPS assumptions
We’ve put in a big effort to find, categorize, and analyze assumptions and guarantees in papers on cyber-physical systems since 2014. Now we’re happy to release the results! Citation: We are also sharing our database of analyzed papers and assumptions.
-
Jordan presents V&V for vision-based systems at ATVA
—
Jordan Peper went all the way to Bengaluru, India, to present our work (in collaboration with UIUC) on unified verification and validation of vision-based autonomy at the International Symposium on Automated Technology for Verification and Analysis (ATVA). Allegedly, this is a hot problem, but the abstraction is quite complex. That’s what it takes — for…
-
IROS showcase: world models, image repair, data cleaning
—
Ivan went all the way to Hangzhou, China, to present several research works on world models, image repair, and data cleaning. Here are the paper citations on which these presentations were based:
-
New preprint: online friction estimation for racing
Our lab pushed out an experimental project on detecting slip and estimating the tire friction from the onboard sensors (lidar & IMU) on RoboRacer (aka F1/10) cars. No fancy models, no sophisticated data collection, no need for post-processing. It turned out pretty accurate!
-
Ivan presents conservative perception abstractions at Allerton
—
Ivan talked about conservative abstractions of perception-driven systems at the University of Illinois Urbana-Champaign in the Allerton Conference. The rumor is that these abstractions are too conservative. Citation:
-
New preprint: how safe will I be given what I saw?
An extension of our modular family of learning-based safety predictors from L4DC 2024, now with transformers and quantization! Citation:
-
New preprint: chain-of-thought confidence with STL
We converted a SAS course project to a workshop paper about how to calibrate the confidence in chain-of-thought reasoning using a temporal logic formula: Stay tuned for more!