How can robots avoid both known and unseen failures in the open world? Take Jenga: a tower can collapse in countless ways when pulling a block. Even without seeing all those failures, our uncertainty-aware latent safety filter (UNISafe) can reliably prevent failures! (1/9)

RT @YilinWu11: I am at RSS this week. Tomorrow I will present our work FOREWARN at Reasoning for Robust manipulation workshop and FM4RoboPl….

RT @kensukenk: I’ll be presenting this work at RSS tomorrow at the Reliable Robotics workshop, and Monday at the Control and Dynamics sessi….

This is the first step of my PhD toward safe and reliable robotics in uncertain, open-world environments. Big thanks to @kensukenk and my advisor @andrea_bajcsy!. PS: Want to feel how safe it is? 🎶 Study With Me — While a Robot Safely Plays Jenga (9/9).

Check out our paper and project website for more on the technical approach, plus extensive results from both simulation and real-world hardware experiments. Project page: ArXiv: (8/9).

We validate our method on a real-world manipulation task—Jenga! Using only raw RGB images, UNISafe proactively corrects unsafe actions with safe, in-distribution actions, blocks risky out-of-distribution actions, while still enabling safe interaction with Jenga blocks. (7/9).

We detect OOD failures by quantifying the epistemic uncertainty of the world model and calibrating an uncertainty threshold via conformal prediction. We then perform HJ reachability analysis in an augmented latent space spanning both the latent and the uncertainty.(6/9)

🔑 Key idea: Use the world model’s epistemic uncertainty as a proxy for identifying unseen failures! By unifying OOD detection with latent-space reachability analysis, UNISafe automatically protects robots from known and unseen failures. (5/9).

When the world model imagines an action sequence entering the purple region, it hallucinates, teleporting the robot into a safe state. This phenomenon makes filters fail to prevent unseen failures, and even known ones, due to overly optimistic imagination of OOD scenarios. (4/9).

Why does this happen? Consider a toy example with a planar vehicle that must avoid two failure regions (gray and purple). The world model is trained on RGB images, but the training data doesn’t include trajectories entering the purple failure region. (3/9)

Our approach builds on latent-space reachability, using world models to analyze safety in complex, vision-based tasks. But when can we trust these world model predictions? World models can hallucinate, overconfidently predicting unsafe out-of-distribution situations as safe.(2/9)