CoRL 2026 · Austin, Texas · Half-Day Workshop

The Science of
Physical AI Safety

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.

View Open Calls Read the Thesis

The Thesis

Why this workshop, now

Robotics is undergoing a paradigm shift — from modular perception–planning–control pipelines to large transformer-based robot foundation models that act directly on the physical world.

Because RFMs inherit the same transformer substrate that frontier AI-safety research has spent years dissecting, the field's tools for interpretability, alignment, and control may already port to robots REF x, y, z.

This workshop brings the two communities into one room to establish the Science of Physical AI Safety — anchored around three questions the field has not yet settled.

The safety case is acute. Failure modes researchers study in LLMs — goal misgeneralization, specification gaming, jailbreaks, deceptive or unintended behavior — become physical harms when the model controls a robot.

When the policy is the robot, an alignment failure is no longer a wrong answer — it is a wrong action in the world.

The day is built around active, collaborative problem-solving: anchor talks and spotlights put concrete context on the table, then breakout rooms reach grounded consensus — with speakers and authors circulating as mentors. Every participant leaves a contributor to a single authored paper.

The Agenda

Three open questions

The workshop is organized around three questions that the field has not yet answered. Each anchors a talk, a breakout room, and a section of the final paper.

01 LLM → RFM

Do AI-safety techniques built for LLMs transfer to robot foundation models?

02 Classical → RFM

Does classical robotics safety transfer to robot foundation models?

03 Evaluation

Are evaluations for RFMs meaningfully different from those for LLMs?

The Programme

Workshop Format

Phase I

Anchor Talks

3 speakers

Three talks on the three key questions

Each anchor speaker frames one of the core questions, putting concrete technical context on the table for the day's collaborative work.

Q1 · LLM → RFM — Do AI-safety techniques built for LLMs transfer to robot foundation models?
Q2 · Classical → RFM — Does classical robotics safety transfer to robot foundation models?
Q3 · Evaluation — Are evaluations for RFMs meaningfully different from those for LLMs?

Phase II

Breakouts

3 rooms

One breakout room per question

Participants opt into a room aligned to one of the three key questions. Each group works to produce:

Their best consensus statement on the question; and
A steelman of the counter-argument against that statement.
Speakers rotate across sub-groups as subject-matter experts.
Spotlight authors rotate across sub-groups as mentors.

Throughout

Spotlights

Interspersed

Spotlight talks & demos

CFP spotlights — short talks from contributed papers.
Safety-failure demos — spotlights from the Demo Call, surfacing RFM failure modes beyond collision.

Outcome

The Paper

Collective

One collaboratively authored paper

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.

Participation

Two ways to contribute

Opening Soon

Both calls will open following workshop acceptance — submission portals and deadlines will be announced here. Accepted submissions are spotlighted on the day, and their authors join the breakout rooms as mentors.

Call for Papers CFP

Work on interpretability, alignment, and control for robot foundation models. Contributed papers are spotlighted and their authors mentor breakout groups.

Submissions open soon · dates TBA