Machine learning · Robotics · Autonomy · Unmanned systems · Applied AI
Systems that pursue goals without being told.
Autelica builds novel products and solutions in machine learning, robotics, autonomy, unmanned systems, and applied AI.
The name
Autelica
Acting of its own accord. The system carries its own loop — sense, decide, act — with nobody holding the controls.
From telos, an end worth reaching. Autonomy is only useful when it is pointed at something.
Machines that hold a goal and choose their own way to it. That is the whole company in one word.
Reconstructing a moving subject in three dimensions needs views far enough apart to triangulate and close enough to agree. The formation is the algorithm. When an obstacle takes a view away, the geometry degrades before the model does — so the aircraft rotate rather than wait.
Domains
Five areas, one discipline: making a system behave correctly on the compute, the timing, and the data it will actually have.
Machine learning
Models built for the constraints they ship into — limited compute, shifting data, and no second chance at inference time.
Robotics
Perception and control that close the loop on real hardware, at real rates, against real noise — without markers, staged lighting, or a room built for the robot.
Autonomy
Planning and decision-making for systems that keep working when the map is wrong and the plan no longer applies.
Unmanned systems
Aerial, ground, and maritime platforms — from bring-up on the bench through trials in the field.
Applied AI
Language and multimodal systems put to work on operational problems. An answer that cannot show where it came from cannot be audited, and anything running unsupervised will eventually be asked.
Vector search returns things that read alike. A graph returns things that are actually connected. Watch a query leave a seed entity and walk the edges — every hop it takes is a step it can later be asked to justify, and the squares it lands on are where the answer came from.
Principles
Autonomy is mostly an engineering discipline, not a modelling one. These are the positions we start from.
Start at the failure
A system that acts on its own is defined by what it does when something breaks. We design the degraded case first — sensor dropped, map wrong, estimate diverging — and work backward to the easy path.
Design for the compute you'll have
The model that wins offline is often the one that cannot run on the platform. Architecture gets chosen against the thermal envelope, the power budget, and the loop rate, not the leaderboard.
Calibrate before you trust
Most failures blamed on an algorithm are extrinsics, clock skew, or a sensor that drifted since morning. Calibration and time synchronisation are part of the system, not part of the setup.
Push the problem into software
Dense rigs and special-purpose infrastructure are the expensive way to buy accuracy. Fewer sensors, cheaper platforms, and a harder reconstruction problem is usually the better trade — and the one that reaches the field.
Simulation is not evidence
Simulation covers what cannot be safely staged, and it is how we reach a first flight without breaking anything. It still proves nothing on its own. Nothing is believed until it has run on the real platform, outdoors, in the wrong weather.
Own the seams
Sensing, models, policy, and the interface an operator actually uses. Each layer can be individually correct and the system still wrong — the interesting failures live at the boundaries, where nobody owns the contract.
Two sensors, one scene, and a few degrees of disagreement nobody measured. Most failures blamed on perception are this: extrinsics, clock skew, or a mount that shifted. The residual is the honest number — until it falls, nothing downstream deserves to be trusted.