AAAI 2027 · February 16–17 · Montréal

Methods in
Physical AI Safety

A tutorial at AAAI 2027 on the methods for interpreting, securing, and steering AI systems that act in the physical world.

Tutorial Forum · February 16–17, 2027
Palais des congrès de Montréal. The setting →
o₁o₂ o₃aₜ OBSERVATION → ACTION TOKENS π(aₜ | o₁:ₜ) policy = RFM a misspecified action becomes → physical harm interpret secure steer
At the conference
Conference
AAAI-27Forty-first annual meeting
Forum
Feb 16–17Tutorials and labs
City
MontréalCanada
Venue
Palais des congrèsIn person
01
The Tutorial

Methods, taught where the system meets the world

Robotics is moving from modular pipelines to general-purpose robot foundation models that act directly on the physical world.

Those models inherit the transformer substrate that frontier safety research already knows how to study. They do not inherit its guarantees. A misspecified objective, a brittle perception stack, or an unmonitored policy can leave the screen and become a physical harm.

The case for treating this as its own problem is set out in Stocking & Häon 2026.

Methods in Physical AI Safety is a tutorial for researchers who need those methods in hand: how to interpret a model that outputs actions, how to secure it against failures that do not stay simulated, and how to steer it once it is in contact with the world.

It is taught for the AAAI community — AI safety, robot learning, and the neighboring fields that will share this problem. The aim is a common technical footing, specific enough to use.

The assigned room and hour will be posted here when AAAI publishes the tutorial schedule. The forum itself runs February 16–17 at the Palais des congrès.

02
The Methods

Three practices

The tutorial is organized around three methods. Each is taught as a practice: the assumption it relies on, the place it breaks on a system that moves, and what to do next.

01 Interpret

Read what a robot foundation model represents, and which interpretability tools survive embodiment.

02 Secure

Test for failures that become physical harm — robustness, monitoring, and evaluation beyond a benchmark score.

03 Steer

Constrain a system whose outputs are actions, where classical control and modern alignment have to meet.

03
The Setting

AAAI 2027, in Montréal

Conference
AAAI-27
Feb 16–23

The forty-first AAAI Conference on Artificial Intelligence

The tutorial is part of AAAI-27, held in Montréal from February 16 to February 23, 2027. Conference page.

Forum
Tutorials
Feb 16–17

Taught during the Tutorial and Lab Forum

AAAI holds tutorials and labs on February 16 and 17, before the main technical program. Attendance is in person. Tutorial registration is separate from the technical program.

Place
Montréal
In person

Palais des congrès de Montréal

The conference meets at the Palais des congrès. Travel and accommodation are posted by AAAI.

Session
To be posted
Room & hour

The schedule will land on this page

AAAI assigns tutorial rooms and times. This page will carry the session once that schedule is published.