Figure says it has delivered more than 350 of its third-generation Figure 03 humanoid robots through its BotQ manufacturing facility, while demonstrating a production rate of one robot per hour after increasing from one per day in under 120 days. The company also described an updated version of its Helix System 0 controller, which combines camera-based perception with the robot’s internal body-state sensing to traverse real-world stairs after training in simulation.

Both developments were reported by Figure in an April 29, 2026 company announcement.

Contents

What changed at BotQ

Figure describes BotQ as having moved from prototype-oriented manufacturing towards production-phase operations. The company says the facility now has dedicated production lines for all critical system modules, supported by custom manufacturing-execution software running across more than 150 networked workstations.

The headline throughput change is from one Figure 03 per day to one per hour. That corresponds to a reported 24-fold increase, achieved in under 120 days. Figure says the one-per-hour figure has been demonstrated and will continue to be optimised, but the announcement does not clarify whether it represents a sustained long-term output rate, a demonstrated production cycle time, or both.

The manufacturing system includes several layers of inspection and testing:

  • Figure says it qualified hundreds of suppliers against incoming-inspection criteria.
  • More than 50 in-process inspection points have been introduced.
  • The company reports an end-of-line first-pass yield above 80%, meaning that more than 80% of robots reportedly pass that stage without rework.
  • Figure reports a 99.3% first-pass yield for its battery line and says more than 500 battery packs have been shipped.
  • It says it has produced more than 9,000 actuators across more than 10 distinct product variants, or SKUs.
  • Each robot undergoes more than 80 functional-verification tests before sign-off.
  • Burn-in testing includes repeated full-body exercises such as squats, shoulder presses and jogging, with cycle counts in the thousands.

These figures describe the company’s manufacturing and quality-control process, but they do not by themselves establish long-term field reliability, operating cost or commercial readiness. First-pass yield measures whether an assembly passes a production stage without rework; it is not the same as a robot’s probability of operating without failure over months or years.

How the controller uses vision

Figure says the earlier version of Helix System 0 used the robot’s joint state, base motion and other proprioceptive information. Proprioception is sensing of a machine’s own body—such as joint positions, motion and internal forces—rather than direct sensing of the outside environment.

That approach can tell the controller what the robot’s body is doing, but it does not directly provide information about the terrain ahead. In the updated System 0, RGB images from cameras in the robot’s head are processed through a stereo model. Stereo vision estimates depth by comparing views from cameras positioned at different locations. The resulting 3D representation is provided to the control policy alongside the robot’s proprioceptive state.

This creates a perception-conditioned whole-body controller. A whole-body controller coordinates multiple joints and body movements together, rather than treating an individual leg, arm or actuator as an isolated system. For stair traversal, the controller must combine information about the visible geometry with the robot’s current balance, motion and joint configuration.

Figure says the policy is trained end-to-end with reinforcement learning in simulation across thousands of procedurally generated and randomised terrains. Reinforcement learning trains a policy—the system that selects actions—using feedback about how well those actions perform in an environment or simulation.

The company says the same network weights that learned to climb generated staircases in simulation were then deployed on a physical robot to traverse stairs in the real world. Figure describes this as “zero-shot” transfer: it says there was no real-world fine-tuning, domain-specific calibration or operator-in-the-loop adjustment, and that the demonstration worked across varying lighting conditions.

Sim-to-real transfer remains difficult because simulated and real environments differ, including in lighting, wear and unexpected obstacles. The announcement does not specify the stereo-model architecture, camera hardware, reinforcement-learning algorithm or computing hardware used by System 0.

Stairs are the first demonstrated use case described in the announcement. Figure presents the architecture as potentially applicable to other behaviours in which the surrounding scene affects the robot’s actions, but the supplied evidence does not establish performance beyond stair traversal.

Why production scale matters

A larger fleet can provide more physical hardware for testing, data collection and deployment. Figure says robots from BotQ are being allocated to internal research and development, data collection, housework efforts, commercial-use-case development and real-world deployment.

The company also says its expanding fleet has helped it perform more diagnostics, build layered software fallback behaviours intended to keep use cases operating or recover from non-critical faults, and investigate lower-frequency edge-case failures. These are issues that may be difficult to identify when only a small number of robots are operating.

Figure reports that it has built an internal Field Service Management system, a Fleet Management System and over-the-air software-update infrastructure. According to the company, these systems support service operations, fleet health and status tracking, software updates, fleet-wide upgrades and recall campaigns.

That operational layer is important for humanoid robots intended to work outside controlled laboratory conditions. Manufacturing more machines is only one part of deployment. Operators also need ways to monitor them, diagnose failures, update software and respond when hardware or software problems occur.

The announcement does not state how many Figure 03 robots are operating at customer sites, in homes or in commercial environments. It also does not provide a consumer price, a complete product specification sheet or general commercial-availability terms.

What to watch next

The most important follow-up evidence will be whether BotQ can sustain the reported one-robot-per-hour rate as production expands, and whether its yields remain stable at higher volumes.

For Helix System 0, more useful evaluation would include stair-traversal success rates, failure and recovery behaviour, speed, supervision requirements and performance across different terrain, lighting and obstacle conditions. Independent testing would help distinguish a controlled demonstration from a broadly reliable capability.

It will also matter whether the perception-conditioned controller generalises from stairs to the wider class of terrain-dependent behaviours suggested by Figure. Finally, customer or residential deployments could clarify how much maintenance and operator support Figure 03 requires in practice.

Sources

  • Figure, “Ramping Figure 03 Production” — https://www.figure.ai/news/ramping-figure-03-production