Research

Physical AI, tested where it has to hold up.

Physical AI is AI that reasons about and acts in the real, physical world rather than only with text. A model that performs on a benchmark and a model that performs on a plant floor are not the same claim. Our research is aimed at the second one.

The team

AI Research for the Real World

AutoAlign's physical AI research team works on the problems that decide whether an autonomous system can be trusted outside a lab: world models that hold up under partial observation, alignment for robotics where a wrong action has physical consequences, and the robustness of physical models in real-world settings.

The work is not separate from the product. Visor is deployed on factory floors, flight lines, field trucks, and mine sites, which means the research programme is tested against the conditions that break systems in practice: degraded sensors, unfamiliar assets, incomplete documentation, and operators who need an answer now rather than a confidence interval.

That reliability focus is where AutoAlign started. NVIDIA integrated AutoAlign into its NeMo Guardrails project, and KPMG deployed and tested AutoAlign within its Trusted AI framework, the same rigour that lets Visor operate in regulated and challenging environments today.

What we research

Real-time control systems

Physical AI representations that create closed-loop and human-in-the-loop decision-making: sensing, computing a response, acting, and repeating, where latency, and deadline guarantees matter.

Robotic alignment

Keeping autonomous and semi-autonomous action aligned with operator intent and safety constraints, in a setting where an incorrect action moves mass rather than emitting text.

Physical model robustness

How multimodal models degrade in real operating conditions (poor light, occlusion, unfamiliar equipment, sensor drift) and what it takes to hold accuracy when they do.

Published

A Sidecar Design for AI Safety

Dan Adamson, Kimi Li, Abhijit Pal, Santhan Kumar Reddy Nareddula, Rahm Hafiz · AutoAlign AI · Published Oct 2024

A holistic architecture for deploying AI in regulated environments, preventing harms, increasing reliability and built as a "safety sidecar" that scales even in low-latency environments as models and usage advance. The design was evaluated across models and integrated into NVIDIA NeMo Guardrails.
Download the whitepaper

What you'll learn

Real-time supervision

A controller architecture that reviews prompts and responses as they happen, correcting issues instead of only flagging them.

Fewer refusals, not more

Alignment controls that reduce error rates while lowering refusal rates: proof that safety and utility do not have to trade off against each other.

Built to scale

A design flexible enough to swap in best-of-breed controllers, including NVIDIA NeMo Guardrails, and to extend to agent frameworks as they mature.

Work with us

Bring the research to your operation.

If you run an operation where reliability is the constraint, we want to hear what you are trying to solve, and where current systems fall short.

Book a demoSpeak with the team