Development project
ATLAS
Advanced Teleoperation Learning for Autonomous Systems
Powered by Digital Holland
ATLAS develops a new AI method that lets industrial robots learn complex tasks by watching skilled operators, including the choices those operators make. By combining choice-aware imitation learning with self-calibrating multi-sensor fusion, robots can handle the variation in objects, layouts and lighting that breaks current automation. TOS Robotics and TU Eindhoven validate the approach in a TRL 4 lab prototype targeting food, manufacturing, logistics and (greenhouse) horticulture.
ATLAS is co-funded by the PPS Innovation Scheme of Topsector ICT (TKI ICT mkb-call 2025).
Robot cells in Dutch SMEs are expensive and rigid. Every product change, new layout or small series requires custom programming, recalibration and scarce specialists, while two thirds of entrepreneurs report ongoing labour shortages. Existing imitation learning approaches copy the demonstrated motion, but ignore the decisions behind it: which object to pick first, in what order, along which route past obstacles. They also break down when a camera shifts or an object is occluded, because sensor setups are hand-tuned and brittle. The result is automation that does not scale to the variable reality of food processing, SME assembly and agri-food production.
ATLAS adds two missing building blocks to imitation learning. First, choice-awareness: the robot learns explicitly which object to grasp first, which sequence to follow and which path to choose when multiple valid options exist. Second, self-calibrating sensor fusion: plug-and-play integration of multiple cameras and depth sensors that stays robust under occlusion or small displacements, without manual recalibration. Operators demonstrate tasks via teleoperation, the system imitates and generalises, and no reprogramming is needed when conditions change. TOS Robotics provides the teleoperation platform, data and industrial integration. TU Eindhoven contributes state-of-the-art machine learning and scientific expertise. The work is delivered as reusable building blocks (code, reference configurations) and an open KPI and benchmark protocol with evaluation scripts.
ATLAS opens a realistic path to affordable, adaptive robot cells for SMEs that today depend on manual labour and migrant workers. It cuts engineering effort and changeover times, raises uptime and shifts people from repetitive physical tasks to quality control and supervision. The shared building blocks and public benchmark let integrators, machine builders, robot suppliers and end-users compare performance and build on common reference implementations, strengthening the Dutch position in smart, human-centred production. Scientific results and the benchmark protocol will be published open access to maximise reuse across the Dutch innovation ecosystem.
Partners
