How to Choose a Stereo Vision Camera for Robotics: A Buyer's Guide

Most stereo vision camera comparison for robotics work comes down to a spec sheet fight over four numbers: baseline, depth range, field of view, and IP rating. Get those four right for your application and everything else, from SDK integration to mounting, tends to fall into place. Get them wrong and you'll spend months fighting noisy point clouds or a camera that can't survive the floor of your facility. This guide walks through what each spec actually controls, when active stereo earns its keep over passive, and how to map camera specs to the robotics tasks you're actually building for.

The four specs that decide fit

Baseline. This is the physical distance between the two lenses in a stereo pair, and it sets the geometry of your depth measurement. A longer baseline gives you better depth resolution at range, because small angular differences between the two camera views translate into larger, more measurable disparities the farther out you look. A shorter baseline does the opposite: it sacrifices long-range precision for stability at close working distances, where a wide baseline would actually struggle to converge. This is why a wrist-mounted camera and a forward-facing AMR sensor use fundamentally different optical designs, not just different housings.

Depth range. Every camera has a minimum working distance where triangulation gives you a usable signal, and a maximum range where noise swamps the measurement. The gap between the two, and where the "optimal" sweet spot sits within it, tells you whether the camera fits your actual working distance or whether you'll be operating it at the edge of its envelope. A camera rated to 20 meters isn't necessarily better for a bin-picking task than one rated to 1 meter; it's just built for a different job.

Field of view (FoV). Horizontal and vertical FoV determine how much of the scene you capture without stitching multiple sensors or accepting blind spots. Wide FoV matters more for AMR navigation, where a robot needs to see obstacles across a broad forward arc, than for a fixed pick station, where a narrower, denser view of a specific work zone is often more useful.

IP rating. This tells you what the housing tolerates, dust ingress and moisture in particular, and it's the spec most often glossed over in early prototyping only to become a blocker at deployment. A camera validated on a lab bench under an IP5X rating behaves differently once it's living on a forklift mast in a warehouse with airborne dust and washdown cycles.

Active-plus-passive stereo versus passive-only

Stereo vision cameras generally fall into two camps. Passive stereo triangulates depth from two camera views using only ambient light, the way human binocular vision works. It draws less power, generates no active illumination that could interfere with other sensors on a crowded robot, and performs consistently regardless of what else is emitting light nearby. The tradeoff is that passive stereo needs textured, feature-rich surfaces to find correspondence between the two views; it struggles on blank walls or low-contrast objects.

Active stereo adds infrared projection to paint texture onto the scene, which is what lets it hold up on featureless surfaces and in low light. Orbbec's Gemini 330 series runs active and passive stereo simultaneously rather than switching between the two, so the camera is drawing on both signal sources at once and picking the reliable one frame to frame. That's a meaningfully different design than a camera that toggles modes, because it means there's no transition lag when lighting conditions change mid-task, which matters for anything moving through mixed indoor and outdoor environments.

For close-range, controlled-lighting work like a wrist-mounted pick camera, passive stereo is often the better fit: lower power draw, no risk of IR cross-talk with a neighboring sensor, and simpler thermal management on a small end-of-arm package. For AMR navigation, obstacle avoidance, or anything that has to hold up in variable lighting from a warehouse dock door to full sun, active-plus-passive stereo is the safer default because it doesn't depend on the scene having enough visual texture to triangulate against.

Matching specs to the application

AMR navigation. This calls for a wide FoV to catch obstacles across a broad forward path, a depth range that comfortably covers several meters so the robot has time to react, and a housing rated for the dust and occasional impact of a warehouse floor. Orbbec's Gemini 335L targets exactly this profile: a 95mm baseline for longer-range precision, a 90° horizontal by 65° vertical FoV, and IP65 protection for industrial environments.

Pick-and-place and bin picking. Arm-mounted cameras live at the opposite end of the spec spectrum. The working distance is short, often under half a meter, so the camera needs a tight minimum working distance and sub-millimeter accuracy at close range rather than long-range reach. Weight matters too, since anything mounted at the wrist adds inertia the arm has to move and control. This is the profile the Gemini 305 was built around: a 4cm minimum working distance, an ideal working range of 7 to 50cm, sub-millimeter depth accuracy at 15cm, and a 68g body that doesn't tax arm dynamics.

Obstacle avoidance. This overlaps with AMR navigation but tends to push harder on reaction time and FoV coverage than on maximum range, since the goal is catching something in the robot's path early enough to act, not mapping the far end of a room. A wide FoV and consistent performance across lighting conditions, rather than raw range, is the priority here.

Robust design

IP rating and housing durability deserve their own line item in any spec comparison because they get decided by where the camera actually lives, not by the robot's software stack. An indoor, climate-controlled work cell can run on IP5X protection without issue. A camera mounted low on an AMR chassis or on a forklift needs IP65 at minimum to handle dust and splash, and anything genuinely outdoors, exposed to rain, temperature swings, or wash-down, needs a purpose-built outdoor rating. Orbbec's Gemini 345Lg is built for that end of the spectrum: IP67-rated, operating from -20°C to 65°C, with a dual-mode depth FoV up to 104° x 87° for the wider coverage outdoor obstacle detection demands.

Two concrete spec comparisons

Two Gemini models side by side show how differently the same underlying stereo technology gets tuned for opposite jobs.

Gemini 305 versus Gemini 335L: the 305 uses an 18mm baseline and passive stereo to hit a 4 to 100cm depth range with sub-millimeter accuracy at 15cm, in a 68g, 42 x 42 x 23mm body rated IP54. The 335L uses a 95mm baseline and active-plus-passive stereo to cover roughly 0.17 to 20+ meters, with an optimal range of 0.5 to 6 meters and depth accuracy of 0.8% at 2m (1.6% at 4m), in an IP65 housing. Same underlying stereo principle, opposite ends of the working envelope.

Gemini 335L versus Gemini 345Lg: both are long-baseline, active-plus-passive stereo cameras aimed at longer-range robotics work, but the 345Lg trades some of the 335L's tighter accuracy spec for a wider operating envelope, IP67 versus IP65, and a temperature range extending down to -20°C versus roughly -10°C for the 335L. If your deployment is a semi-controlled indoor or light-industrial setting, the 335L's spec profile is usually sufficient. If it's a genuinely outdoor environment with real weather, the 345Lg's rating exists for a reason.

Integration and SDK considerations

Spec sheets only tell half the story. All three cameras run on Orbbec's open-source SDK v2, with native ROS and ROS2 wrappers and cross-platform support across Windows, Ubuntu, macOS, and Android. For teams already running a RealSense-based pipeline, it's worth knowing that Orbbec's stereo line generally offers a shorter minimum working distance and a broader working envelope than comparable RealSense cameras, along with in-camera depth processing that keeps computation off the host, which matters when you're running perception on an embedded compute module rather than a full workstation.

A few common questions

Do I need active stereo if my robot only operates indoors under consistent lighting? Not necessarily. If the environment has enough visual texture, textured floors, boxes, shelving, passive stereo can hold up fine and saves power. The risk is blank walls, glossy floors, or glass, where passive stereo loses correspondence points. If any part of your workspace has those conditions, active-plus-passive is the safer choice.

How much does baseline actually matter for close-range picking? More than people expect. A long-baseline camera designed for meters-scale range often can't converge reliably at the centimeter-scale distances a pick task requires. That's why the 305's short baseline isn't a compromise for arm-mounted work, it's the correct design choice.

What's the practical difference between IP54 and IP65 for a warehouse AMR? IP54 protects against limited dust ingress and splashing water from any direction, fine for a clean indoor cell. IP65 is fully dust-tight and withstands low-pressure water jets, which is closer to what a warehouse floor with forklift traffic, dust, and occasional spills actually demands.

The fastest way to waste a robotics vision budget is picking a camera based on a marketing headline rather than the four specs that actually govern whether it works in your environment. Start with baseline and depth range for your working distance, FoV for your coverage need, and IP rating for where the camera physically lives, then let the application dictate active versus passive. The camera that wins on paper for someone else's use case is rarely the one that's right for yours.