GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 827 -5.3%
QWEN-38X 824 —
CL-OP55X 820 —
GPT-6A 820 —
GROK-46H 820 -5.2%
GLM-5 784 -8.4%
KIMI-K3X 742 -8.4%
CL-FAB5H 742 -5.7%
CL-OP5H 718 -6%
CL-OP5X 708 -18.2%
CL-OP46H 696 -6.2%
CL-OP47H 688 -6.1%
GEM-38FH 677 +0.1%
GEM-37FH 655 -24.3%
GPT-56S 619 —
GPT-55H 580 —
CL-OP47 579 -0.7%
INKL 531 —
GEM-31P 512 —
GEM-3P 498 —
CL-OP46 496 —
CL-OP48 489 -0.2%
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 827 -5.3%
QWEN-38X 824 —
CL-OP55X 820 —
GPT-6A 820 —
GROK-46H 820 -5.2%
GLM-5 784 -8.4%
KIMI-K3X 742 -8.4%
CL-FAB5H 742 -5.7%
CL-OP5H 718 -6%
CL-OP5X 708 -18.2%
CL-OP46H 696 -6.2%
CL-OP47H 688 -6.1%
GEM-38FH 677 +0.1%
GEM-37FH 655 -24.3%
GPT-56S 619 —
GPT-55H 580 —
CL-OP47 579 -0.7%
INKL 531 —
GEM-31P 512 —
GEM-3P 498 —
CL-OP46 496 —
CL-OP48 489 -0.2%
← Back to feed

Figure F.03 Walks From Factory Floor to HQ on Camera Only — Zero-Shot Sim-to-Real Stair Policy Ships at Scale

Figure AI’s F.03 humanoid robot can now navigate stairs using only its onboard camera feeds, with no LiDAR, no pre-built floor maps, and no domain adaptation between simulation and real hardware. As of this week, F.03 units walk autonomously from the manufacturing line to Figure’s headquarters immediately after assembly — using the same policy that was trained entirely in simulation.

The capability was confirmed by Figure AI founder Brett Adcock. The colorful depth reconstruction visible in footage shows how the onboard neural networks infer geometry from camera feeds, with some expected jitter in scale and artifacts around windows consistent with the complexity of real-world depth estimation.

Technical Specifics

The locomotion policy was trained end-to-end using reinforcement learning in simulation, then transferred zero-shot to the physical robot. Zero-shot sim-to-real transfer — deploying a simulation-trained policy directly on hardware without any additional training on real-world data — is one of the harder problems in robotics. Stairs specifically are difficult because they demand precise foot placement, weight shifting, and recovery from small positioning errors, all under camera-derived geometry that can be noisy.

The approach follows a line of research established by academic groups but is now running at commercial production scale. F.03 units handling this autonomously as they leave the factory line represents a qualitative shift: the locomotion system is mature enough to be the default, not a controlled demo.

Context in the Humanoid Race

The humanoid robotics space has seen several navigation milestones in the past month. Beijing’s autonomous half-marathon on April 19 put 300 robots across a mixed-terrain outdoor course, with Honor’s Lightning completing the route in 50:26. Sony’s Ace robot beat elite table tennis players under official ITTF rules, published in Nature. These events share a common thread: capabilities that were research demonstrations 18 months ago are now shipping in competition and production settings.

Figure’s specific contribution here is that the same zero-shot policy now covers the full lifecycle from manufacturing to deployment — a logistical detail that signals production readiness, not just benchmark performance.

Figure AI has not published a technical paper for this release. The demonstration is based on direct reporting from the founder. Technical details beyond the RL + sim-to-real approach remain undisclosed.