Shenzhen-based Opticalix has launched a vision-perception suite built specifically for embodied AI, betting that the bottleneck for robots leaving the lab is no longer brains but eyes. The package centres on AI-driven binocular stereo vision and two modules, Libra 1000 and Libra 3000, that the company says stabilise depth recovery in exactly the scenes where classical stereo vision falls apart: transparent glass, polished metal, dark objects and near-featureless walls.
The pitch is straightforward. As multimodal models and vision-language-action (VLA) systems mature, robots are getting better at deciding what to do. But the moment they enter a home, factory or shop, their usefulness is capped by how reliably they perceive the physical world. Autonomous navigation, fine grasping and human interaction all sit on top of a layer of three-dimensional visual understanding. If that layer is thin, the whole stack wobbles.
Why classical stereo vision stalls
Opticalix argues traditional stereo matching, which leans on local texture, breaks down on the surfaces robots actually meet. Transparent and reflective materials produce missing or noisy depth; low-texture walls confuse correspondence; black objects vanish. The result is flawed point clouds that corrupt obstacle avoidance, grasping and navigation.
Its answer is to fold an in-house stereo-matching algorithm into the perception pipeline so the system reasons about scene structure and object relationships, not just pixel contrast. On transparent and weak-texture surfaces, the company claims it recovers continuous, detail-rich depth and rebuilds millimetre-level geometry on tiny objects.
One vision system, many jobs
The suite uses what Opticalix calls full-domain generalised perception: a single hardware platform handles depth, SLAM localisation and object detection, and supports over-the-air upgrades. That contrasts with the traditional approach of bolting on separate sensors for each function. Two products anchor the line: Libra 1000, a 2 cm-baseline module aimed at close-range manipulation on wrists and grippers, with a fisheye option pushing field of view past 180 degrees; and Libra 3000, a 7 cm-baseline head-mounted module for mid- to long-range avoidance, mapping and spatial understanding.
The company also says it trimmed the compute budget, so more silicon is left for VLA inference and localisation. For Western readers tracking the robotics supply chain, the signal is that Shenzhen’s component makers are moving up from selling cameras to selling perception systems, the layer that actually decides whether a robot works in the wild.
Read the original report (LeiPhone)
Translated and adapted from LeiPhone (leiphone.com).