Machine vision is undergoing a structural transition. For decades, digital video has relied primarily on frame-based cameras that capture full images at fixed intervals, creating large amounts of redundant data. Research teams are now exploring neuromorphic engineering, modeled on biological nervous systems, to address the latency and power limitations of conventional cameras.
Spanish-Basque technology center Ikerlan is pursuing this approach through the BEGI project. BEGI is a single-lens, passive, neuromorphic 4D sensor that mimics an insect eye to capture 3D depth and track motion using just one optic.
Ikerlan claimed its technology could significantly improve data processing and power consumption. But to move from theory to real-world products, the team must overcome hardware limitations, find early markets, and eventually design its own chips.
Commercial possibilities and prototype
To build the BEGI sensor, the team places a microlens array on top of a dynamic vision sensor. Unlike conventional cameras that capture still frames, each pixel in this sensor operates independently and responds only when it detects a change in incoming light. This creates a steady stream of visual events that show movement almost instantly and use very little energy.
View All The microlens array shapes the incoming light so the sensor can determine its direction. Using a method called feedforward stereopsis, the system uses this extra information to quickly calculate 3D shapes and motion, all without needing a second camera.
However, building a new class of sensor from the ground up requires significant capital and time. To rapidly demonstrate the BEGI concept, the Ikerlan team chose to build its initial prototype using existing commercial hardware rather than immediately developing a custom chip.
“We have to have an event sensor, a neuromorphic sensor,” Xabier Iturbe, project lead at Ikerlan, said in an exclusive interview with EE Times. “We have long-term plans to develop it ourselves in collaboration with other Spanish partners such as the [Consejo Superior de Investigaciones Científicas] CSIC. But in the short term, to demonstrate the concept, which is what we have done, we need to go fast, and we need to use established technology to work on something existing.”
The team chose the IMX636, a 0.9-megapixel dynamic vision sensor from Prophesee, built on Sony technology. Iturbe said Prophesee is a well-known European company with strong funding, but the main reason for choosing the sensor was its high pixel count. When the project started, he said the IMX636 offered the highest resolution in event-based vision.
Solving resolution constraints
Resolution remains the main challenge for the current BEGI sensor. To measure depth, the system needs multiple pixels to capture the same spot from different angles, so it requires a large number of pixels for the calculations to work well.
The 0.9-megapixel sensor limits how far and wide the current BEGI prototype can see. Iturbe said that, in theory, a perfect resolution could provide a 2-meter range with a 40-degree field of view, or up to 10 meters with a narrower 20-degree field of view. In practice, however, the prototype operates at a resolution of about 0.2 pixels, limiting its range to less than 1 meter and its field of view between 20 and 40 degrees.
“That delimits a lot of what we can do with these physical prototypes that we have today,” Iturbe said. “Mainly, as you can see, we have to go to short-range applications.”
Applications in aerospace and industry
Because the current hardware has a short range, Ikerlan is focusing on industrial and aerospace applications where short-range operation is important. In factories, robots and machines often operate within a meter of their targets, making the BEGI prototype well suited to these environments.
Aerospace is another good fit, especially for the precise maneuvering required to dock satellites, Iturbe said. Existing aerospace technologies perform well for long distances, but the final step of bringing two spacecraft together requires very accurate, real-time sensing.
“What we know from the space people we have talked to is that the long distance is solved with technologies they have,” Iturbe noted. “But the short distance, when you are very close, and you have to give the last thruster adjustments to connect and dock well, it seems there are challenges. Everyone says this is an interesting technology for that.”
The design of the neuromorphic system also makes it well suited for spacecraft, where power and thermal budgets are tightly constrained. “We capture the data that provide the most information to the application,” Iturbe said. “The derivative of functioning that way is that we only spend energy and time to process data that are truly valuable.”
Push for a custom ASIC
To move BEGI from a specialized industrial tool to a key part of the wider physical AI market, including drones, self-driving systems, and advanced robots, the sensor needs a much bigger range and field of view. To get there, the team must stop relying on commercial sensors and build its own custom chip.
Switching to a full-frame, high-resolution custom sensor could allow BEGI to see up to 16 meters deep and cover a 120-degree field of view. This would put it in direct competition with established stereovision systems such as Intel RealSense. BEGI would offer comparable spatial awareness while adding benefits such as microsecond response times, a dynamic range exceeding 120 decibels, and simpler single-lens hardware that avoids the calibration challenges of two-camera setups.
“When we make the leap to a full-frame sensor, a large-size sensor with large resolution, which today would mean designing and manufacturing it ourselves… the metrics skyrocket,” Iturbe told EE Times. “We are getting into robotics, we are getting into drones, we are getting into navigation, manipulation. We are getting into the whole revolution that is expected to happen related to physical AI.”
Ecosystem integration
BEGI’s commercialization strategy goes beyond just the sensor hardware. Ikerlan is also working to make the sensor data usable across many types of computing systems.
With agencies such as NASA and the European Space Agency starting to approve open-standard RISC-V chips for space, and private companies developing their own chips, the BEGI team is ensuring its technology works with any hardware.
“Our message is very clear: Anyone who wants to process our data is welcome,” Iturbe emphasized. “We don’t care if it’s RISC-V, we don’t care if it’s Arm… If someone wants to put a GPU there, no problem.”
Ikerlan’s primary goal is to bring this technology to the broader industry. The team is helping establish a new European group focused on neuromorphic design while creating a dedicated semiconductor design center to help others adopt the technology.
Although building a full-frame ASIC remains a tough engineering challenge, the BEGI sensor’s underlying architecture marks a transition away from bulky, data-intensive cameras toward more efficient, nature-inspired machine vision.
Read also:
Neuromorphic Spikes Unify Control and Decision Making In this episode of Brains and Machines, professor Rodolphe Sepulchre, a control theorist from the University of Cambridge, discusses the inspiration he took from studying biological neurons, why both discrete and continuous behaviors are inherent to how they work, and why building neurons is often easier than simulating them.
Indian Researchers Develop Molecular Memristor for Neuromorphic Computing IISc Bangalore Researchers have developed a molecular memristor with 14-bit analog resolution and 4.1 TOPS/W energy efficiency.