Marco Pavone
Stanford & NVIDIA
Bio ↗
Convening the AI safety and robot learning communities to establish a shared science — and a shared research agenda — for the interpretability, alignment, and control of robot foundation models.
Robotics is undergoing a paradigm shift — from modular perception–planning–control pipelines to general-purpose robot foundation models that act directly on the physical world.
Because RFMs inherit the same transformer/diffusion substrate that frontier AI-safety research has spent years dissecting, the field's tools for interpretability, alignment, and control may already transfer to robots Häon et al. 2025 Buurmeijer et al. 2026 Swann et al. 2026 Robey et al. 2026 Joseph et al. 2026 Jeong et al. 2026.
The safety case for Physical AI is acute Stocking & Häon 2026. Failure modes researchers study in LLMs — goal misgeneralization, specification gaming, jailbreaks, deceptive or unintended behavior Amodei et al. 2016 Hendrycks et al. 2021 Sharkey et al. 2025 — could become physical harms when the model controls a robot.
Yet the two communities best equipped to address this rarely meet: AI safety seldom touches physical hardware, and robot learning seldom engages frontier interpretability and alignment.
This workshop convenes them to establish the Science of Physical AI Safety — organized around three questions the field has not yet settled: what AI-safety tools transfer from LLMs, what guarantees carry over from classical robotics, and whether evaluating robot foundation models demands fundamentally new methods.
The event is built for active, collaborative work rather than passive listening: anchor talks and spotlights set the technical context, then breakout rooms drive toward grounded consensus, with speakers circulating as mentors. Every participant leaves a contributor to a single, collectively authored paper that names the field's open research problems and sets a shared research agenda.
The workshop is organized around three questions that the field has not yet answered. Each anchors talks, a breakout room, and a section of the final paper.
Do AI-safety techniques built for LLMs transfer to robot foundation models?
Does classical robotics safety transfer to robot foundation models?
Do RFMs require evaluation techniques that are meaningfully different from both LLMs and classical robotics?
Anchor speakers frame each of the core questions, putting concrete technical context on the table for the day's collaborative work.
Participants opt into a room aligned to one of the three key questions. Each group works to produce:
Breakout artifacts are aggregated into a single paper that names the field's open research problems and sets a shared research agenda. Every participant is a contributor.
Submission portals and deadlines will be announced soon. Accepted submissions are presented on the day, with a selection invited to give spotlight talks.
We welcome and value diverse perspectives — and strongly encourage submissions across all career stages, genders, ethnicities, geographies, and institutional backgrounds. This is a real and meaningful commitment, not a formality: capturing the full range of viewpoints is this workshop's central purpose.
Work on interpretability, alignment, and control for robot foundation models. All accepted papers are presented as posters, with a selection invited to give spotlight talks.
Visual media plus optional short analysis of an RFM safety failure beyond collision. Selected demos are spotlighted on the day.
