If you put a robot vacuum in a living room, it will meticulously map the baseboards. Move a single armchair, however, and the machine may spin in confused circles, trapped by its own digital blueprint.

These robots are prisoners of geometry. Lacking adaptability, the machines can be completely disoriented by a simple change, like moving a chair, and halted in their progress.

To solve this issue, Chinese scientists from Northwestern Polytechnical University have developed a new “bio-inspired cognitive navigation” framework.

This framework aims to give robots “spatial intelligence” — the innate ability humans and animals have to navigate unfamiliar, dynamic environments using internal spatial awareness and past experiences.

Bio-inspired framework

Animals easily find their way through new places using flexible mental maps, memory, and smart planning — all while using barely any energy.

However, robots lack an integrated cognitive architecture. Old-school robot navigation is like walking with a paper map. The robot follows a step-by-step process and cannot handle any route that isn’t already drawn out.

This reliance on high-energy processing and rigid programming fundamentally limits operation in unpredictable, real-world settings.

Led by Professor Guo Bin, the research team drew inspiration from how mice navigate mazes. Instead of forcing a machine to memorize a room pixel by pixel, they are reportedly teaching it to think like a maze-running mouse.

Mice recognize key landmarks, form abstract memories, and build “cognitive maps” to flexibly apply knowledge to new situations.

The underlying problem with modern robotics is a lack of abstraction. When a mouse explores an unfamiliar barn, it does not record every grain of wood. It identifies key landmarks, files away compressed memories of the layout, and builds a fluid cognitive map in its brain. When conditions change, the mouse does not panic. It adapts, reasons, and takes a shortcut.

“Memory plays an active role in navigation by compressing experience into reusable knowledge and reconstructing it on demand,” the researchers noted in the study paper. 

To replicate this biological advantage, Guo’s team designed a three-pillar framework based on dynamic landmark recognition, experiential memory, and hierarchical decision-making.

Neuromorphic hardware

Pairing this cognitive software with brain-inspired “neuromorphic” hardware, the team has created a highly efficient system layout. 

These specialized processors mimic biological neurons by activating only when detecting changes in sensory input. This approach eliminates the constant power drain typical of standard computing, clearing the path for a new generation of ultra-low-power, agile, and autonomous machines.

“Neuromorphic sensors and processors, which activate only in response to input changes, provide an energy-efficient substrate that can meet the latency demands of on-board robotic deployment,” the team noted. 

The new framework lets robots pinpoint location, predict their surroundings, and use past experiences in new places. This allows them to plan routes flexibly and shifts them from just following orders to making their own smart decisions.

If a machine is to successfully navigate a burning building during a rescue operation, or safely assist an elderly person in a cluttered home, it cannot rely on a pre-programmed layout. It must expect the unexpected.

This development could enable robots to stop blindly calculating and start thinking by teaching machines to distill experiences into reusable knowledge.

The team is currently collaborating with various organizations to transition this technology into practical, real-world deployment.

The findings were published in the journal Nature Reviews Electrical Engineering on May 22.