The R2 started as an internal tool: something to carry parts between our own assembly and burn-in halls. It became a product the day a customer watched it navigate a crowded warehouse floor and asked when they could buy one. Here is what is inside.
Mapping once, using it forever
The R2 builds a semantic map on its first walkthrough — not just geometry, but what the obstacles are: pallets that move, racks that do not, doorways that need clearance. Replanning against a known map costs a fraction of online SLAM, and that saving is what buys battery life. When the floor changes, the map updates incrementally instead of starting over.
Local inference, by design
Every perception decision runs on the robot itself. A cloud round-trip adds latency you can feel in a corridor full of people, and a dropped connection turns a helpful robot into an obstacle. The vision stack is quantised and tuned for the R2’s NPU, and the planning model runs alongside it — the same agent runtime as our cloud products, compiled down.
Battery economics
Eight hours of autonomy from a 48 V pack is a budget, not a spec sheet number. Roughly 40% goes to drivetrain, 35% to compute, 15% to payload, and 10% is reserve. The biggest win came from duty-cycling perception: full-resolution vision while moving, low-power tracking while stationary. Duty-cycling alone added 90 minutes per shift.
What we would do differently
Hot-swappable batteries, from day one. Customers love the eight-hour figure until their shift pattern is twelve hours, and a screwdriver is not an acceptable answer to that. It is on the R2s.