A clip of Figure's F.03 humanoid that spread widely on 1 August as an autonomous ladder climb is a two-hour timelapse of the robot repeatedly walking up and down stairs. Figure founder Brett Adcock's own description says the tests are helping move the company's robots "closer to fully autonomous systems" - a phrase the reposts converted into a claim that this run was autonomous. Both changes happened in the same headline.
Key facts
- The artifact is a timelapse of roughly two hours of repeated stair traversal by the F.03 humanoid, not a ladder climb.
- Figure's own wording is "closer to fully autonomous systems," not a claim that the run was unassisted.
- Figure does state elsewhere that the videos listed in its Helix 02 write-up are fully autonomous and not teleoperated - a claim scoped to those specific clips.
- Primary sources: Adcock's original post and Figure's Helix 02 page.
Two words changed and the story changed with them. "Stairs" became "ladder," which upgrades a locomotion problem into a whole-body climbing problem. And "closer to fully autonomous" became "autonomously," which converts a direction of travel into an arrival.
This is worth being precise about, because the underlying evidence is genuinely interesting. Repeated stair traversal over a long test interval is a real durability and control signal. Stairs are a nastier problem than they look: each step is a discrete contact event where the robot must place a foot on an edge it cannot fully see, absorb the impact, and keep its centre of mass moving without the recovery margin that flat ground provides. Doing it for two hours without the run being abandoned says something about thermal behaviour, joint wear, controller stability and repeatability that a slick thirty-second demo cannot.
What a timelapse cannot establish is equally specific: whether an operator intervened, whether the two hours were continuous, what the failure rate was, whether the robot recovered from stumbles or was reset, whether it generalises to a different staircase, and whether foot placement was conditioned on perception or on a known geometry. Time compression is not evidence of deceptive editing - a two-hour video is unwatchable and nobody expects a raw feed - but it does remove precisely the information that would settle the autonomy question.
The mechanism underneath is documented, and it is the part worth reading. Figure describes Helix 02 as a three-layer hierarchy. A high-level semantic layer sets goals. A visuomotor layer converts onboard vision, touch and proprioception into whole-body joint targets at 200 times per second. A lower learned controller handles balance and contact at 1,000 times per second. The point of that stack for stairs is that balance, limb contact and body pose stay coupled instead of the robot switching between a walking controller and a manipulation controller. It is the difference between a person who plans a route and adjusts their footing continuously, and one who has to stop planning in order to look down.
The hardware side is converging on the same thesis. AMD's Ryzen AI Embedded X100 puts CPU orchestration, GPU perception, neural inference and shared memory into a single rugged embedded part explicitly positioned for physical AI and autonomous robotics. The argument is not that a chip confers autonomy; it is that a robot whose camera processing, model inference, planning and real-time control fight across separate devices and memory pools will lose milliseconds it cannot spare. AMD's comparisons against competing modules are vendor claims and should be read as such.
The useful research counterpoint arrived the same week. Caltech's PAC-MAN paper, submitted 30 July, treats robot safety as a perception problem rather than a control problem. During training the system knows the exact geometry of an incoming ball relative to every robot link and is rewarded for maintaining clearance. At deployment that privileged information is deliberately taken away: the policy sees only a head-mounted, segmentation-masked depth view plus its own joint states. The finding is the interesting bit - a lightweight per-link safety barrier survives that transition, while a stronger joint-space safety filter only wins when it is fed an accurate ball state the deployed camera stack cannot actually supply. In a 20-throw hardware test on a Unitree G1, the robot took one impact and did not fall.
That is exactly why the Figure framing matters. A safety mechanism that assumes information the deployed robot does not have will fail in the field regardless of how good the demo looked. Google's Gemini Robotics 2 release set a better disclosure standard here, openly reporting that several fine manipulation tasks succeed only around a third to under half the time.
The honest caveat cuts both ways. Nothing here suggests Figure misrepresented anything - the company's own language is careful, and the exaggeration happened downstream. But careful language in a post that is destined to be screenshotted is a weak defence, and companies shipping viral robot footage now have an obvious remedy available: state the autonomy conditions and the failure boundary in the clip itself.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)