Seen vs. Followed
Do robotic foundation models act on text that appears in the scene?
Edge AI engineer working on AI safety. Creating monitoring methods and evaluations for robotics and agents.
- Who I am
- Edge AI engineer working on Physical AI safety for Generalist Robotics. Track record of deploying computer vision projects in production. Researcher on Physical AI Safety through hackathons, workshop papers, and public code.
- What I do
- I build monitoring methods and evaluations to improve AI oversight deployed in robotics and agent systems.
- How I can help you
- If you are building simulations to monitor robotic foundation models or agentic systems creating policies for robotic movements, I can help ensure AI safety test cases are monitored through system one auditing methods, baseline ML methods, and better evaluation designs.
Writing
- Seen vs. Followed: A Protocol for Measuring Contextual Awareness of In-Scene Text in Vision-Language-Action Models
Zachary Gunderson
SPAIS workshop @ CoRL 2026 · submitted · Oct 1, 2026The submitted version covers the earlier LIBERO-Goal experiments. The DROID follow-up is on the project page.
- How robotics could speed up the intelligence explosion Physical AI Safety · Substack · Oct 2026
- Layterms AI Safety · Layterms · Sep 2026
A nonpartisan walk-through museum of AI.
Projects
I Spy Agent
Small decision models as black-box monitors for covert agent behaviour: sabotage hidden in files (LinuxArena) and collusion through covert table talk (NARCBench-Stego).
Stance drift in multi-turn agent conversations and steering
I used steering vectors to try and improve at run time the models' actions in the tutoring setting. Ultimately I succeeded in my examples but had poor generalizability.