Spatial Memory
Persistent Spatial Memory
A map tells a robot where things are. Spatial memory tells a robot what it has learned from moving through the world.
The Problem with Maps
Why robots need spatial memory
Traditional maps are static snapshots. They show geometry at a moment in time, but they do not capture what a robot learns by moving through an environment repeatedly.
In the real world, environments change. Doors open and close. Obstacles appear and disappear. Elevators have schedules. Human traffic follows patterns. Ground conditions vary with weather and time.
Spatial memory captures all of this: not just the structure of a space, but the history of what has happened there, what works, what fails, and what changes over time.
Memory Architecture
Four layers of spatial memory
Beaver Robotics builds memory across four interconnected layers that together form a persistent world state.
Spatial Memory
Spaces, structures, object locations, semantic landmarks, route history, and scene versions. This layer captures the physical and semantic structure of the environment as the robot observes it over time.
Temporal Memory
Human flow, door states, elevator patterns, dynamic obstacles, and historical changes. This layer captures how the environment changes over time — hourly, daily, weekly, and seasonally.
Affordance Memory
Where a robot can move, avoid, pass, climb, dock, hand off, interact, or collaborate. This layer encodes what actions are possible in each part of the environment, specific to each robot form.
Action Memory
What happened after an action, which strategy worked, where failure occurred, and how the world changed. This layer connects robot actions to outcomes, enabling learning from real-world experience.
World Model
From static maps to persistent world memory
A 2D map captures geometry. A 3D map captures structure. But neither captures time, experience, or possibility.
A 4D world model adds the dimension of time and experience. It records not just what a space looks like, but how it behaves, what has happened there, and what a robot can expect when it returns.
Every route a robot takes, every obstacle it encounters, every failure it experiences, and every successful passage it completes adds to this growing memory.
The result is a world model that gets richer and more accurate with every deployment — creating a compounding intelligence advantage.
True intelligence is not reinterpreting every frame from zero. It is building reusable memory through long-term interaction with the world.