Skip to content
ArchXS

Practice

AI that moves things, not only documents

Learned policies replaced hand-written trajectories, and that changed which question decides a robotics project. Not which robot to buy, but whether the loop from demonstration to deployed policy closes, and what each stage has to prove before the next one is paid for.

AI that moves things, not only documents
Fig. 01The loop from a teleoperated demonstration to a deployed policy, closing in days, on a staged path where every gate is a place at which stopping is a decision rather than a failure.

How we look at it

Until recently a robot cell was a program per product: a trajectory written by hand, one task per cell, an integrator project measured in years, and nothing reusable the moment the product changed. What changed is not that the arms got better but that policies became transferable, a vision-language-action model fine-tuned on a hundred teleoperated demonstrations, on open weights, with transfer across embodiments, so the unit of work is a dataset rather than a program. That moves the risk elsewhere. The question is no longer whether the arm can execute the motion, it is whether the loop closes: whether a failure seen on the floor turns into new demonstrations and a better policy within a week. And the second question is not robotic at all, it is where the plant's IT ends and its OT begins, because that seam is where these projects actually stall.

What we have built

The industrial layer underneath is not a plan, it runs: production planning to ISA-95 across four levels with a CP-SAT constraint solver as the decision-maker, a rolling horizon and plan-freezing zones; acquisition over OPC UA to IEC 62541 through our own bridge into the event bus; MQTT with Sparkplug B; telemetry in TimescaleDB. On top of that we designed a robotics pilot for a mid-size manufacturer as a feasibility map rather than a proposal. The stack is named and open: Isaac Sim and Isaac Lab for physics and domain randomisation, LeRobot for imitation learning and versioned datasets, ROS2 with MoveIt 2 for collision-aware motion, SAM2 and FoundationPose for perception and grasp poses, a VLA policy fine-tuned with rank-32 LoRA and served as a 30 Hz action stream over ONNX and a ROS2 bridge, failures recorded to rosbag2 and replayed into the next round of demonstrations. So is what gets measured: task success rate, first-try grasp, the sim-to-real gap, recovery outside the training distribution, policy latency. The path has four phases, each with a technical readiness criterion: simulation only, proving the policy converges on the target task at no hardware risk; a desktop research arm, proving it transfers and the sim-to-real gap is measurable; a small cobot in a controlled zone, proving the success rate holds on a single product; a production cell, proving the policy generalises to new products without retraining. One boundary stated plainly: what we have delivered in production is the planning and IT/OT layer, not a robot cell. The robotics work is a designed pilot with its gates written down before the first purchase order, which is the only form in which we are willing to begin one.

What follows

Hardware commitment is deferred until something has been proven in simulation, and each later phase buys a specific piece of evidence instead of a promise from an integrator. Whoever stops at a gate stops holding a simulation environment, a demonstration dataset and a measurement, not a cell that performs one task. And because the demonstrations and the policy are the artifacts, the second product costs a fraction of the first, which is the only version of this that pays back.