Autonomy that survives contact with the real world.

The site captured to a measured model before the first robot arrives: aisles, doors, docks, clearances and the places people walk. The robots’ maps are built from it, and when the layout changes the change is captured and the maps updated, instead of the fleet being re-taught by hand.

• The site captured before the first robot arrives
• Clearances and pedestrian routes measured, not assumed
• Robot maps built from the measured model
• Layout changes captured and the maps updated

A warehouse is to receive its first autonomous trucks in March. In February the building is captured in an afternoon, and the model shows two aisles narrower than the drawing and a fire door that opens into the planned route. The route is moved before a single robot is delivered. In May a new rack line goes in, and one capture pass updates the maps overnight.

Task-level intent decomposed into something a platform can execute, supervised continuously, and interrupted the moment the world stops matching the plan. Accountability is recorded per action.

• Task intent decomposed into executable steps
• Supervised continuously and interrupted on divergence
• Per action: what was intended, observed and done, and who authorised the class

At 08:00 a supervisor sets a task: move rack C to bay 4 by 09:00. It is broken into four steps a pallet mover can execute, and supervised as they run. At 08:20 bay 4 turns out to be occupied. The world no longer matches the plan, so the running step holds instead of pressing on, and "place at bay 6" goes to the supervisor, who approves it at 08:23. It is done at 08:55, and every action is on the record: what was intended, what was observed, what was done and who authorised it.

One board for every robot in a site, whatever its maker: jobs allocated across the fleet by what each robot can carry and where it is, crossings and narrow aisles reserved so robots never meet head-on, and charging fitted into the gaps. People see one fleet, not several.

• Robots from different makers on one board
• Jobs allocated by what each can carry and where it is
• Crossings and narrow aisles reserved in turn
• Charging fitted into the gaps between jobs

A distribution centre runs pallet movers from one maker and tote robots from another. At the crossing by the dock they used to stop and wait for each other. Now the crossing is reserved in turn: a pallet mover with a full load goes first, two tote robots hold for eleven seconds, and a robot low on charge is sent to charge in the gap before the evening peak. The shift supervisor sees one queue of jobs for all of them.

The guidance stack tuned for the final stretch: degraded sensing, unexpected geometry, and recovery from being wrong rather than an assumption of never being wrong.

• Tuned for doorways, loading bays and thresholds
• Holds up under degraded sensing and unexpected geometry
• Recovers from being wrong instead of assuming it never is

Crossing the open floor, the robot follows its planned path at full speed. In the doorway to the loading bay someone has left a pallet, and the sensors see only its near face. The estimate admits it is unsure; the robot slows to a crawl, sees the pallet’s top, finds a 1.2 m gap and threads it. It docks at the bay to the centimetre, and nobody had to step in. The last hundred metres are where most autonomy stops.

A perception platform that reconstructs the space it is moving through as it moves, at a fidelity good enough to plan against. The same reconstruction stack that produces survey-grade models offline, constrained to run at the edge inside a motion budget.

• Reconstructs the space while moving through it
• Fidelity good enough to plan against, not merely avoid
• The survey stack, inside an edge motion budget

A platform enters a storage level it has never seen, with only an old plan to go on. As it moves, it reconstructs walls, openings and objects well enough to plan against, not merely to avoid. The plan says doorway A is 1.0 m wide; the reconstruction measures 0.82 m, too narrow for the platform, so the route is re-planned through opening B before it gets there. It leaves with a model good enough for the next platform to plan on.

The humanoid platform doing repetitive work in spaces built for people: moving totes between conveyor, rack and bench at the heights and reaches the line was designed for. It works more slowly than a person and does not need to be fast; it needs to keep going through the shift. Anything it cannot handle goes to the person supervising the cell.

• Repetitive moves in a line built for people, left unchanged
• Totes between conveyor, rack and bench at human heights
• Steady through the shift rather than fast
• Anything it cannot handle goes to the cell’s supervisor

On a line that was never designed for automation, the humanoid moves totes from the end of a conveyor to a rack across the aisle, all night. At 02:40 a tote arrives split open; the platform sets it aside instead of guessing, and flags it. The supervisor clears it in two minutes, and the line keeps its night shift without anyone doing the lifting.

The quadrupedal platform walking a plant’s inspection round several times a day: reading analogue gauges, taking thermal images of pumps and motors, listening for leaks, and comparing each reading with the rounds before. Anything outside its band goes to the control room with the image, and the operator decides.

• Gauges read, pumps and motors imaged, leaks listened for
• Every reading compared with the rounds before
• Stairs, grating and uneven ground crossed
• Anything outside its band sent to the control room; the operator decides

On the 03:00 round the platform reads forty gauges and images twelve pumps. Pump 7 runs 14 °C warmer than on any round this week, while its gauge still reads normal. The thermal image and the trend go to the control room; the operator checks it on the morning walk, finds a failing bearing, and the repair is planned before it can stop the line.

The quadrupedal platform walking a site every week and capturing it to a measured model, compared with the design. Walls, openings and services that are out of place are flagged while they can still be moved, and progress is measured rather than estimated.

• A weekly walk of the site, captured to a measured model
• Compared with the design, element by element
• Out-of-place work flagged while it can still be moved
• Progress measured, not estimated

On the weekly walk of level 3, the capture shows a service opening in a new wall 180 mm from where the design puts it, before the ducts arrive. The site engineer sees the overlay the same afternoon, the opening is moved, and Monday’s walk confirms it. Without the walk, the mistake would have surfaced when the ductwork did not fit.

When a robot meets something it cannot resolve, it stops safely and asks. One operator on the board sees the robot’s view and its options, chooses one, and the robot carries on; the answer becomes a rule or a map update, so the same question is not asked twice. Every intervention is on the record.

• A robot that cannot resolve a situation stops safely and asks
• One operator answers many robots from one board
• The answer becomes a rule or a map update
• Every intervention on the record

At 11:20 a delivery robot finds a cleaning sign in its corridor and cannot tell whether it may pass. It stops and raises a hand. The operator, who is watching twenty robots, sees its camera view and chooses to pass on the left, and the sign is marked temporary on the map. Two other robots reach the same corridor in the next hour and pass without asking.