Designing physical AI for human unpredictability

Ali Shafti on why progress in physical AI will depend on robots that can understand context, intent and the people they work alongside.

“For physical AI to succeed, it must operate naturally in real environments alongside human teammates. Those environments are rarely tidy or predictable. They are messy, dynamic and full of human behaviour that doesn’t follow a script.

Humans deal with this effortlessly because we learn to build shared mental models with those around us, from an early age. We understand where we are, what we’re trying to do, what each of us is capable of, and what the other person is likely to do next. That shared context is what makes collaboration feel natural.

For robots, none of that is inherent – it must be built from the ground up.

We approach this by capturing and maintaining context on four blocks: environment, task, capability and intent. If a system can track those in real time, interaction becomes far more robust – even when confronted with human unpredictability.

From there, different levels of interaction emerge. A robot may simply need to recognise people and move safely around them. In other cases, it interprets natural instructions through speech and gesture. More advanced systems anticipate needs and assist proactively. At the most capable end, human and robot work together on shared tasks, adapting as conditions change.

This is what we’re building at CC. Our physical AI systems use a sense–understand–support loop that connects perception and context with reasoning and action. They fuse multimodal sensory inputs, to maintain a structured understanding of the human, task and environment, and update that context continuously.

That shared understanding allows the system to respond appropriately as situations evolve – adapting to new scenarios while retaining safety and reliability.

When that works, robots stop feeling like tools we operate and start feeling like teammates we can work alongside.”