AI Vision & visual intelligence
Detection, segmentation, depth estimation, tracking and pose — trained on your footage and validated against the failure modes that matter: glare, dust, occlusion, motion blur, night operation.
Division 01
YTK Robotics takes machines from sensing to acting. We build the visual intelligence, the radio link, the onboard inference and the autonomy layer that lets a humanoid, an industrial arm or a drone operate in the real world — where light is bad, the network drops out and the cost of being wrong is physical.
A detector that scores well on a clean dataset tells you almost nothing about a machine working at 45 °C with a smeared lens and a vibration problem. So we engineer the whole chain: sensor selection, calibration, timing, compute budget, comms and the fallback behaviour when any of it degrades.
That is why the division holds five disciplines in one team rather than handing a spec between vendors. Vision, radio and inference are coupled problems; treating them separately is how robotics programmes stall.
Fig. 01 — perception and latency, as instrumented in the field
Each capability stands alone as an engagement, and all five are designed to interlock. Most programmes start with one and end up touching three.
Detection, segmentation, depth estimation, tracking and pose — trained on your footage and validated against the failure modes that matter: glare, dust, occlusion, motion blur, night operation.
Multi-rotor and hybrid platforms, flight control tuning, mission autonomy, precision positioning and payload integration for inspection, survey and patrol.
Link budget analysis, antenna selection, mesh and long-range telemetry, spectrum planning and interference mitigation for contested airspace and industrial sites.
Mission-level autonomy: task decomposition, tool and payload control, replanning on failure, and supervision that keeps a human accountable for consequential actions.
Smaller models running locally on Jetson-class and embedded accelerators — quantised, pruned and scheduled to fit real thermal and power envelopes without a cloud round trip.
Humanoids, industrial robots and drones as complete systems — perception, compute, radio and autonomy brought together and tested as one machine.
Every platform we build follows the same five-layer spine. It keeps the failure boundaries explicit: perception never commands an actuator directly, and the radio layer is treated as unreliable by design.
For a machine that moves, latency is not a performance metric — it is the distance travelled before it reacts. We shrink the model until it fits the compute envelope on the device, and keep the network for coordination rather than control.
Typical trade at the deployment stage: a modest accuracy delta in exchange for an order-of-magnitude reduction in reaction time, plus operation that continues when the link does not.
Indicative figures from lab benchmarks. Actual budgets are set per platform.
Not a list of markets — a list of conditions. If your deployment looks like one of these, the engineering is already close.
Assets that are tall, hot, pressurised or hard to reach. Close-range visual capture, thermal overlay and defect flagging, with the drone holding position in wind.
Corridors, sites and structures mapped repeatedly to a consistent georeferenced baseline, so change between surveys is measurable rather than anecdotal.
Autonomous routes, anomaly detection and human escalation — designed so an operator can verify any alert quickly enough to act on it.
Indoor positioning without reliable GNSS, mixed human and machine traffic, docking and handover, and pick-and-place where misidentification is expensive.
Whole-body perception, grasp planning and safe physical interaction in spaces built for people — with torque and proximity limits enforced below the planner.
Send the operating conditions. We will tell you what is feasible and what is not.