How 3D vision technology is making collaborative robots safer on the factory floor

Collaborative robots have earned their place on the plant floor by working alongside people instead of behind fencing, but that arrangement only holds up if the robot can actually perceive the person next to it. 3D vision technology for collaborative robot safety has become the practical answer to that problem, giving cobots the depth data they need to detect a human presence, hold a safe operating zone, and adjust behavior in real time as conditions change around them.

Why 2D sensing falls short for collaboration

Traditional safety systems, light curtains, laser scanners, and fixed-zone photoelectric sensors were built for an era when robots and people didn’t share space. They’re good at telling a system that something crossed a plane. They’re not good at telling it what crossed the plane, how fast it’s moving, or how far away it actually is in three dimensions. For a robot arm meant to slow down near a technician’s hand but keep working near a stationary tote, that distinction matters.

This is the gap that ISO/TS 15066 was written to close. The specification extends ISO 10218’s baseline industrial robot safety requirements to define four collaborative operating modes: safety-rated monitored stop, hand guiding, power and force limiting, and speed and separation monitoring (SSM). SSM is the mode most dependent on real-time spatial awareness, since it requires the robot to continuously recalculate a protective separation distance and adjust speed or stop before that distance is violated. That calculation needs volumetric data, not a binary trip signal, which is why depth cameras have become central to how integrators design collaborative workcells.

What depth data adds to the safety case

A depth camera returns a point cloud, a three-dimensional map of everything in the robot’s field of view, updated many times per second. That lets a safety system track the actual distance between a human operator and the nearest point on the robot, not just whether a zone has been breached. As the operator moves closer, the robot can reduce speed proportionally instead of stopping outright, which keeps the cell productive without compromising the separation distance the standard requires.

Stereo vision cameras are well suited to this task because they combine active and passive stereo simultaneously rather than switching between the two, maintaining consistent depth output whether the work cell is under bright overhead lighting or in a dim corner of the plant. Orbbec’s Gemini 330 series, for example, uses this simultaneous active-plus-passive approach alongside the company’s MX6800 depth-engine ASIC, which handles depth processing and depth-to-color alignment inside the camera itself. For integrators, that in-camera processing matters because it reduces the compute burden on the robot controller or edge PC, which is often already handling motion planning, perception for pick-and-place, and other real-time workloads.

Accuracy at working distance is the other variable that determines whether a vision-based safety system holds up in practice. The Gemini 335, for instance, is rated at 1.5 percent depth accuracy or better at 2 meters, and the longer-baseline Gemini 335L improves on that at the same range while extending usable coverage further out. Those numbers matter for SSM because the protective separation distance calculation is only as reliable as the depth measurement feeding it. An undersized margin built on noisy depth data either forces overly conservative speed reductions that hurt throughput, or worse, an undersized margin that doesn’t actually protect the operator.

Robust design for the actual plant floor

Cameras deployed for cobot safety don’t sit on a lab bench. They’re mounted near moving equipment, exposed to dust, vibration, and temperature swings, and expected to keep working through a full production shift. IP-rated housings matter here as much as the depth specification does. Orbbec’s Gemini 335 carries an IP5X rating, and the longer-range 335L steps up to IP65, giving integrators options depending on how exposed the mounting location is. Matching the housing rating to the environment is as much a part of the safety design as the sensing modality itself.

Real-time response as the environment changes

A cell that’s safe at startup isn’t necessarily safe an hour later, once pallets have shifted, a cart has been rolled into the aisle, or a second operator has entered the space. This is where continuous 3D sensing outperforms a safety system designed around a static configuration. Because the camera is reading the full scene rather than a fixed detection plane, the robot’s safety logic can respond to new obstacles and new people as they appear, without requiring a technician to remap the zone.

Orbbec has built out its Gemini stereo vision line specifically with this kind of dynamic industrial deployment in mind, including validation with NVIDIA’s Isaac Perceptor and Jetson Thor platforms for robotics workflows, along with ROS and ROS 2 support that lets integrators bring depth-based safety logic into existing robot control stacks without building a custom perception pipeline from scratch.

As collaborative robots take on more tasks that used to require full separation from people, the sensing layer underneath them is doing more of the safety work than it used to. Depth cameras give integrators a way to meet that requirement with data, not just a trip wire, which is likely why 3D vision is becoming a standard line item in cobot safety design rather than an optional add-on.