RealSense Moves Robot Vision Closer to the Edge

The D585 Pro combines 3D depth, onboard processing and edge AI in one sensor for humanoids, AMRs and industrial arms.

RealSense introduced the D585 Pro at Automate 2026 in Chicago on June 18, alongside Perception Studio, a new software layer for “Physical AI” applications. The relevant point is not just that another depth camera has arrived. RealSense is trying to put 3D perception, image processing and some edge AI inference into one robot-ready module, reducing the amount of work that has to be handled by the robot’s main computer.

According to the official announcement, the D585 Pro is built around a new Gen 5 system-on-chip, a compact processor that combines several compute functions in a single component. The company claims more than twice the depth quality of previous RealSense generations and 2.5 times better close-range performance than competing solutions. The target uses are broad: humanoid robots, autonomous mobile robots, collaborative arms, industrial inspection systems and human-robot interaction.

The specifications show the direction of travel: a 120 by 100 degree field of view, 60 frames per second at 1280 by 960, IP65 protection and detection below 15 centimeters at full resolution. That short-range number matters for manipulation. A robot that picks up a tool, inspects a part or works close to a shelf needs reliable perception near its own hands or gripper, exactly where many depth cameras become less useful.

The larger issue is robot architecture. By moving more processing into the camera, RealSense wants to simplify perception stacks and reduce dependence on host computing resources. For integrators, that can matter as much as raw resolution: fewer external synchronization steps, fewer host-side workloads and software features delivered through an SDK can shorten trial cycles, especially when several prototypes need to be compared quickly in a lab. It also gives robotics teams a clearer baseline when moving from a demo cell to a pilot fleet. It also frames perception as an upgradeable platform, rather than a fixed peripheral bolted onto a robot.

That fits a wider robotics trend: pushing models and sensing closer to the physical task, in order to improve latency, robustness and integration. A camera that can handle more of the pipeline locally may also reduce bandwidth pressure and make perception easier to reproduce across fleets. The claim will still need to be tested in real deployments, but the signal is clear. Robotics competition is not only about better AI models or better motors. It is also about sensors that can turn messy physical environments into usable data quickly enough for machines to act.