What we're working on
Working notes, not finished papers
Overview of safety and compliance for autonomous machines
A landscape survey and gap analysis of safety and compliance across autonomous vehicles, robotic arms, and humanoid robots, and why the deterministic safety paradigm breaks down as intelligence moves to the center of the system.
Read more→Different bets on autonomy: how training and data choices decide what can go wrong
There is no agreed path to full autonomy, so companies are betting on very different ways to train their machines, from end-to-end imitation on fleet data to simulation and teleoperated demonstrations. Each bet carries a distinct safety signature. This work maps the bets and asks, for each, which failures it reduces, which it hides, and which assurance methods it actually admits.
Read more→Humanoids among people: the safety impact of moving general-purpose robots into shared human spaces
Humanoid robots are leaving controlled environments and entering warehouses, hospitals, and homes, where they share space with people who never trained to work beside a machine. This work characterizes the hazard classes a deployed humanoid introduces and asks whether today's robot safety standards were built for a body like this.
Read more→Spatial semantics inference: the perception a machine needs to be safe around people
Most physical-AI safety failures begin as perception failures. This work studies a safety perception layer we propose an autonomous machine needs to handle safety-critical edge cases, by completing occluded 3D geometry and inferring physics from partial sensor input.
Read more→Safety world models: predicting harm before a machine acts
A world model is a learned predictor of how a scene will unfold. Most are built to improve capability. This work studies the world model built for safety: one that forecasts the futures in which a person gets hurt and lets a machine rule those actions out before it commits to them.
Read more→A real-time safety framework: enforcing safety inside the control loop, not after it
Most safety work happens offline, in standards, test suites, and verification done before deployment. But a learned machine acting among people fails at runtime, in situations no test anticipated. This work studies the runtime layer that monitors perception, prediction, and action while the machine operates, and intervenes within the control loop's time budget.
Read more→Propose a topic
Let's work on safety for Physical AI together.