The visual pathway for robotics.
The visual pathway for robotics.
The visual pathway for robotics.
Why it matters
Why it matters
Why it matters
Robots have eyes. They do not have a visual pathway.
The human visual system does not process every pixel. The retina takes in signals from over a hundred million photoreceptors and sends about a million onward, and what it sends is not an image but a selection of what matters. That is how a person can play soccer, land an aircraft, or perform surgery on about twenty watts.
Robots have been built the other way around. Every pixel gets the same bits, and the model is left to work out what mattered. When that falls short, the remedy has been more cameras and a larger model. That works in a data center. On a robot, it runs, but slower than the cameras and at the cost of the battery.
What a robot can afford to see is the bottleneck on its intelligence. We are building the visual pathway that breaks it: the layer between the camera and the model that decides what a robot actually sees.
Select: Keep only what bears on the task, before anything spends compute on it.
Understand: Turn what remains into objects, motion, and relationships.
Deliver: Get that understanding to whatever is deciding, in time to act.
The human visual system does not process every pixel. The retina takes in signals from over a hundred million photoreceptors and sends about a million onward, and what it sends is not an image but a selection of what matters. That is how a person can play soccer, land an aircraft, or perform surgery on about twenty watts.
Robots have been built the other way around. Every pixel gets the same bits, and the model is left to work out what mattered. When that falls short, the remedy has been more cameras and a larger model. That works in a data center. On a robot, it runs, but slower than the cameras and at the cost of the battery.
What a robot can afford to see is the bottleneck on its intelligence. We are building the visual pathway that breaks it: the layer between the camera and the model that decides what a robot actually sees.
Select: Keep only what bears on the task, before anything spends compute on it.
Understand: Turn what remains into objects, motion, and relationships.
Deliver: Get that understanding to whatever is deciding, in time to act.
What we do
What we do
What we do
We build the visual pathway in three parts.
We build the visual pathway in three parts.
Compression
We turn pixels into conceptual understanding.
Our pathway keeps only the visual information a robot needs and represents objects, motion, and relationships rather than raw frames.
Compression
We turn pixels into conceptual understanding.
Our pathway keeps only the visual information a robot needs and represents objects, motion, and relationships rather than raw frames.
Compression
We turn pixels into conceptual understanding.
Our pathway keeps only the visual information a robot needs and represents objects, motion, and relationships rather than raw frames.
Evaluation
We analyze thousands of hours of video to find the cases that matter.
Because the pathway understands video rather than storing it, it can search an entire dataset for the rare moments where robots fail.
Evaluation
We analyze thousands of hours of video to find the cases that matter.
Because the pathway understands video rather than storing it, it can search an entire dataset for the rare moments where robots fail.
Evaluation
We analyze thousands of hours of video to find the cases that matter.
Because the pathway understands video rather than storing it, it can search an entire dataset for the rare moments where robots fail.
Hardware–software co-design
Real-time understanding at a fraction of the power.
We shape the model and hardware around each other, so the pathway runs at the speed of the cameras and within a robot's power budget, without giving up accuracy.
Hardware–software co-design
Real-time understanding at a fraction of the power.
We shape the model and hardware around each other, so the pathway runs at the speed of the cameras and within a robot's power budget, without giving up accuracy.
Hardware–software co-design
Real-time understanding at a fraction of the power.
We shape the model and hardware around each other, so the pathway runs at the speed of the cameras and within a robot's power budget, without giving up accuracy.
TRUSTED BY
