Robotics
Foveated vision
Definition
Foveated vision processes selected image regions at higher spatial detail while retaining lower-detail context over a wider field of view. A robot can choose or predict the high-resolution region according to the task instead of processing every pixel uniformly.
Also known as: Foveal vision, Visual foveation
Updated
Preserve context and inspect a selected region
Uniformly downsampling a wide camera view can erase small grasp features, while processing the full image at its original resolution increases memory and computation. Foveated vision keeps a broad contextual view and allocates greater detail to one or more selected regions.
The region can come from a fixed gaze model, a detector, a human cue, or a learned policy. When the robot chooses where to look based on its task and current observation, foveation is a form of active perception. The sensing action changes which visual detail is available to the next decision.
A policy can predict its next visual crop
In the 2026 FoveaVLA preprint, Chen and colleagues train a vision-language-action model to predict an interaction region alongside each action chunk. At the next policy query, the model receives original-resolution crops from external cameras together with the full views.
The authors report real-robot task-suite success rates of 70.8% for their single-view configuration and 77.1% for their dual-view configuration, compared with 52.1% for their wrist-camera baseline. These figures describe their post-trained models, DROID-based data, camera arrangements, tasks, and evaluation protocol. They do not establish that a foveated input will outperform a wrist camera in every workspace.
The useful detail can still be missed
Foveation differs from ordinary cropping because region selection is part of the sensing or policy design, and the broad view is normally retained. It also differs from optical zoom: a digital crop cannot recover detail that the sensor never captured.
Selection errors can exclude the contact point, a small obstacle, or a second object needed for context. Rapid motion can make a previously chosen crop stale, and multiple simultaneous interactions compete for limited high-resolution regions. Camera calibration, latency, crop resolution, and the selection policy must all be evaluated. Foveated vision changes the allocation of visual computation; it does not by itself solve recognition, depth estimation, or control.
Sources
Related terms
Active perception
Active perception is perception in which a robot chooses motions or interactions partly to obtain more useful observations. Examples include moving a camera to reveal an occluded object or touching an object to reduce uncertainty about its state.
Vision-language-action model
A vision-language-action model is an AI model that uses visual observations and language instructions to produce actions for a robot. It connects what a robot sees and what it is asked to do with outputs that a robot controller can execute.
Visual servoing
Visual servoing uses visual measurements inside a feedback loop to control robot motion. The controller updates movement to reduce an error defined from image features or visually estimated pose.
Open-vocabulary object detection
Open-vocabulary object detection locates objects in an image and associates them with text-defined categories beyond a fixed closed list used during detector training. It combines spatial detection with vision-language representations that can be queried using class names or phrases.