Real buildings do not behave like labs or warehouses.
Most robots perform well in controlled environments. Buildings where people live are different: narrow hallways, elevators, residents who pause and turn unpredictably, building-specific workflows, and social norms that actually matter for task completion.
The failure mode is rarely dramatic. It is supervision. The moment a human has to watch closely, step in, or work around the robot, autonomy is already breaking down. At scale, that friction becomes the product.
The Behavioral AI layer for robots doing work around people.
Generalist brain companies build the motor stack — navigation and manipulation across embodiments. But making robots useful to people requires a judgment layer above that: understanding what people need, learning per-person preferences, and translating behavioral signal into commands. That layer does not exist at meaningful scale. We are building it through live deployments, building by building, task by task.
The system runs across three layers: See — spatial search that identifies what matters across floors, rooms, and dynamic indoor spaces. Think — spatial reasoning that infers, adapts, and responds. Act — spatial execution that turns perception into real-world action: navigation, handoffs, retries, and human fallback when needed.
Each deployment feeds all three layers. Every completed task strengthens the models. The system becomes more reliable with every building it operates in.
Lily is a prototype, not the product. The behavioral AI layer running above the motor stack is the product. Hardware-agnostic by design.