Figure says its BotQ manufacturing facility now builds one Figure 03 humanoid every hour, up from one per day roughly four months ago. The Sunnyvale company posted the update on its production blog on April 29, claiming over 350 third-generation robots have rolled off the line. CEO Brett Adcock added on Figure's X account that 55 robots will ship this week.
About those numbers
The 24x throughput jump is the headline figure, and it sounds dramatic until you remember the starting point was one robot a day. Going from a near-prototype rate to an early production rate is exactly the kind of curve that flatters its own slope. Still, 350 units is real metal on real shelves, and the 55-this-week claim is the sort of number that gets disproven quickly if it's wrong.
Yields tell a more interesting story. Figure puts end-of-line first-pass yield at "over 80% and improving weekly," which is a hedge worth reading carefully. Battery yield, by contrast, sits at 99.3% across more than 500 packs shipped. That gap between the body and the battery is where most of the manufacturing pain is hiding. Actuators have crossed 9,000 units across more than 10 distinct SKUs, which is a lot of variety to be optimizing simultaneously.
Each robot then has to survive over 80 functional tests before sign-off, including burn-in sessions of thousands of squats, shoulder presses, and jogging cycles. The rationale: catch early-cycle failures before the robot leaves the building. The unstated assumption is that enough robots are still failing in burn-in to make this worth advertising.
Helix sees the floor now
Buried in the same announcement is a control-system update that's arguably the more technically interesting story. The company's whole-body controller, called Helix System 0, used to run purely on proprioception. Joint angles, body motion, no vision. Walking flat ground was fine; stairs required hand-tuned mode switches and an operator nudging things along.
The new version pipes RGB images from head-mounted cameras through a stereo model, lifts them into a 3D scene representation, and feeds that to the policy alongside the usual joint-state inputs. Training happens entirely in simulation across thousands of randomized terrains using reinforcement learning. The same weights, Figure claims, transfer to the real robot zero-shot. No real-world fine-tuning, no domain-specific calibration, no operator in the loop. If that holds up under outside scrutiny, it's the sort of sim-to-real result the field has been chasing for a long time. The demonstration is stair traversal. The architecture, the company says, is general.
The fleet as feedback loop
Figure's framing is that every robot off the line doubles as a data-collection engine. More units in the field means more edge-case failures surfacing, which means more training signal flowing back into Helix. It's a tidy loop, and one that only works if the units actually get deployed somewhere useful. Some go to internal R&D, some to commercial pilots, some to the company's home-robotics push that arrived with the Figure 03 launch last fall.
BotQ is targeting 12,000 units per year on its first-generation line, with a stated goal of 100,000 robots over four years. That second number is doing a lot of work in investor decks. Hitting 12,000 in 2026 means roughly 33 a day, sustained. The current 55-a-week pace, if it holds, gets nowhere close.
The next data point is whether Figure hits the 55-robots-a-week claim sustainably and whether the 80% yield actually improves the way Adcock says it will. The company has put numbers on the table that are easy to falsify, which is more than most humanoid robotics outfits are willing to do.
