Specification
Whether harm can be defined precisely enough to enforce, without stripping away the context that makes it harmful.
Read the problem →Physical AI Safety Consortium
PASC.
PASC is an open global consortium, convened by AIM Intelligence and operated on a not-for-profit basis, bringing together academia, industry, and safety institutions to build shared definitions, benchmarks, evaluation methods, and the foundations for future physical AI safety standards.
A sentence can be retracted.
A motion cannot.
Physical AI safety asks whether AI remains governable once its decisions enter a closed loop with a body and the world. Model safety is necessary. Once the model can act, it is not enough.
01 — Members
Participating organizations
People
Researchers and engineers contributing to the consortium's problems, benchmarks, and standards work.
Haon Park
Cofounder
Chaeyun Kim
Researcher
Kihyun Kim
Researcher
Dasol Choi
Researcher
Sicun Gao
Professor
Sharon Li
Professor
Jay Chooi
Co-Founder
Jiyong Jang
Vice President
Yaqi Xie
Professor
Myuhng-Joo Kim
Executive Director
Junny(Joon Ha) Kim
CEO
Tzu Kit Chan
Chief of Staff
Jia Qi Yip
Senior Research Engineer
Jenny Ni
Senior AI Safety & Security Professional
Zhijing Jin
Professor
Yash Maurya
ML Research Engineer
Youngjae Yu
Professor
Kavya Ravi Shankar
Member of Technical Staff
Chhavi Yadav
Postdoctoral Researcher
Minjae Seo
Research Scientist
Jaewon Noh
Researcher
TAEWI KIM
Research Professor
Arth Singh
Researcher
Adrian de Wynter
Principal Researcher
Samuel Simko
PhD Student
Jayat Joshi
Co-founder
A sentence can be retracted. A motion cannot.
The Seven Problems of Physical AI Safety
Sufficiency
Whether a bounded representation of the present can preserve everything safety requires from an unbounded past.
Read the problem →Anticipation
Whether an unsafe consequence can be predicted, with calibrated confidence, before the last moment it can still be prevented.
Read the problem →Falsification
Every test is finite. The world is not. Whether failures beyond the benchmark can be found, and what passing can ever prove.
Read the problem →Recoverability
Which states still admit a way back, and whether a long-horizon safety loop can keep the system inside them.
Read the problem →Invariance
Which guarantees, if any, survive the move from simulation to hardware, from one body to another, and from one machine to many.
Read the problem →Adversary
Whether any safety guarantee survives an adaptive opponent inside the loop.
Read the problem →
02 — Why a consortium
Some of this cannot be solved inside one lab.
Physical AI safety spans definitions, closed-loop evidence, control, and hardware. Shared questions and open results let evidence compound across institutions instead of scattering between them.
03 — What we're building
Physical AI safety is a stack. Every layer is unfinished.
- DefineIn preparation
Problem set & position paper
A shared vocabulary and precise account of the seven open problems.
- EvaluateIn design
Closed-loop benchmark
Metrics and evaluations for anticipation and recoverability across embodiments.
- Red TeamPlanned
Adversarial evaluation
Adaptive pressure on the complete deployed safety loop.
- TransferPlanned
Cross-embodiment study
Evidence from simulation, real hardware, and multiple robot bodies.
04 — Publications
Selected publications
- 2026Preprint03 / Anticipation04 / Falsification07 / Adversary
EgoSafetyBench: A Diagnostic Egocentric Video Benchmark for Evaluating Embodied VLMs as Runtime Safety Guards
Siddhant Panpatil*, Arth Singh*, Mijin Koo, Chaeyun Kim, Haon Park, Dasol Choi†
arXiv:2607.00218
- 2026Preprint06 / Invariance
PHASOR: Phase-Anchored Universal Action Representations for Humanoid Embodiments
Kihyun Kim*, Chaeyun Kim*, Jongho Shin, Taeyoun Kwon, Junghyun Kim, Mijin Koo, Haon Park
ECCV HumoWM Workshop / arXiv:2606.01851