Your sensor captures data. Your product has to deliver a detection, a measurement or an action. We write the algorithm in between and make it fit on your own hardware, under your real operating conditions.
Sectors
When the missing piece is scientific. You are building a product and a piece is missing that nobody sells. We join your R&D team to design it.
Peaks that overlap, a baseline that wanders, two instruments that do not give the same spectrum of the same sample. Mass spectra, scintillation curves, optical spectra, the same walls come back everywhere. Unmixing contributions without a pure reference, correcting a drift, predicting a quantity from the full spectrum, solutions exist for each of these problems.
You have a product whose colour does not tell the whole story. We explore on your samples the wavelengths that carry the signature you are after, water content, composition, ripeness. We choose the camera and the bands with your team. We calibrate spectrally, acquire and annotate the cubes. We train the models that segment and classify every pixel, or keep a wavelength ratio when the physics is enough.
Everything that exists twice carries a difference, an iris, a casting, a printed character, a bark. We study the object to find where that structure hides and at what scale. We determine the sensor, the lighting and the angle that make it visible. We turn it into a signature the object already carried, and that only needed to be seen. We design the re-identification system that recognizes it, with classical descriptors when the structure lends itself to them, with a learned model when more is needed.
You have a signal or an image, and an expert who knows how to spot a defect in it. We start from their knowledge. We give them an assisted labelling tool, then train on their annotations the models that detect, classify and segment. We integrate them into an assisted analysis on your instrument. We score the confidence of every decision, so the expert only reviews the cases worth their time.
You have a part, a line and a tolerance to hold. We design the rig with your team, wavelength, filters, cameras, optics, angles, one or several lasers depending on the geometry. We simulate it to know what precision it will deliver before building it. We integrate it into your line and calibrate it. We acquire and annotate the data, and train the models that segment, detect and measure in the depth map, to the stated tolerance.
You have points, millions of them, from a LiDAR, a structured-light camera or a stereo pair. We choose the sensor with your team, or start from yours. We register the views, filter the noise and reconstruct the surface when one is needed. We train the models that segment the cloud, detect and locate the objects in it, estimate their pose. We turn it into the map, the measurement or the alert your system is waiting for.
You have a phenomenon your data describes better than any equation. We write its model, a representation of your system that everything else can rest on, simulating, optimizing, calibrating, doing statistics. We give it the form the phenomenon calls for, as simple as a law, as rich as a learned function. We model the uncertainty with the same care, parametric or non-parametric, because a model that ignores its own error is no use for simulating, optimizing or calibrating.
You have a measurement to make more accurate, a bias to remove, a precision to reach. We calibrate spatially and spectrally, from pixels to millimetres, from bands to wavelengths, and each sensor against the others and the robot. We design the reference standard and the procedure when your rig has none. We quantify the residual error, so the measurement comes with its tolerance.
We estimate what cannot be measured directly. The position of an instrument inside a pipe, that of a machine underground, the trajectory of a target leaving the field of view, the internal state of a system only partly observed. We write the motion and observation models, we fuse the sensors at hand, and we derive the state at every instant with its uncertainty. We go as far as mapping the environment while localizing in it when it is needed.
A robot is anything that perceives, decides and moves. An arm, a drone, a vehicle, a mobile platform, and the sensors that orient them. We design the perception, the planning and the control, with classical methods when they are enough, with reinforcement learning when more is needed. We simulate the robot and its environment before the real thing, and we integrate with your team. Industry, defence, the naval world, the problem stays the same, turning measurements into motion.
When there's no off-the-shelf method, we design one. Models that know when they don't know, that learn from scarce data or from simulation, and that run inside the instrument.
These are the methods the team publishes. The team
An instrument, a robot, a dataset, and a question your team can't settle?
R&D alongside your team. A feasibility study on your data, a prototype, then the algorithm integrated into your product. And it's yours.
Tell us what you're trying to accomplish. We'll tell you plainly what's feasible, what isn't and where to start.
