F.02 Decommission

TL;DR

Figure retires its F.02 humanoid fleet by having the robots autonomously jump into a 75-ton electric-arc furnace at a steel foundry in Imatra, Finland — the only facility willing to accept lithium-ion-laden robots. The team trained a jump policy on-site in San Jose using airbags and stunt-artist reference motions, then deployed it in simulation to a foundry the robots had never visited, requiring six successful jumps into a bucket of molten steel within 24 hours. The post is primarily a product-marketing/history piece but it incidentally documents an unusually extreme sim-to-real deployment: policies ran successfully in high-heat, high-electromagnetic-field conditions that disabled surrounding camera and electronics gear. Resulting steel is being machined into limited-edition commemorative artifacts.

Key claims

Figure trained F.02 robots to jump accurately into a target bucket using airbag practice at HQ, stunt-artist motion as a reference, and sim-trained AI policies [§3].
The policy transferred zero-shot to a never-before-visited foundry environment and succeeded within a 24-hour window across six melt attempts, each with only a ~20-minute window before the molten steel crusted over [§4].
Robot AI policies ran reliably despite extreme heat and electromagnetic fields from the 75-ton electric-arc furnace, while surrounding camera equipment and other electronics failed in the same environment [§4].
No metal foundry in the US or Mexico would accept lithium-ion-powered robots into their equipment; the final deployment required an operator in Imatra, Finland [§3].
Most of the F.02 fleet has been destroyed; the recovered metal is being machined into a small-run commemorative artifact series [§5].

Method

Not a research publication — the post gives operational narrative, not benchmarks. The disclosed training pipeline: (1) at Figure’s San Jose campus, set up large airbags on the second floor of a building and iterate on robot jumping; (2) use motion-capture / reference from professional stunt artists as the target trajectory; (3) train a new policy in simulation that targets accurate landing in a specified bucket location; (4) deploy to the Finland foundry with no prior real-site training. Robust-execution evidence is anecdotal (six-of-six success in a narrow window, electronics-hostile environment) rather than quantified.

Results

Six melt windows attempted, all successful enough to produce usable steel bars (exact per-jump accuracy not reported).
Policies continued executing when camera equipment and other electronics in the furnace area failed — but no quantitative robustness characterization is provided.
A limited-edition commemorative artifact series is being produced from the resulting steel.

Why it’s interesting

For a wiki that mostly tracks research artifacts, this one is unusual: it’s a product-launch / corporate-storytelling piece whose only research-adjacent content is a single sim-to-real stunt deployment. Still, two things stand out. First, the training recipe — airbags + stunt-artist reference motion + sim policy → zero-shot deployment to an unseen high-heat site — is a concrete anecdotal data point for the “sim is sufficient for high-DOF stunts if the reference motion is good enough” claim that comes up around work like Scaling Behavior Foundation Model for Humanoid Robots (ScaleBFM’s PPO motion-tracking recipe scales to large humanoid behaviors from retargeted mocap). Second, the whole exercise is downstream of Figure’s broader data-and-fleet strategy around Helix: F.02 is being retired because the F.03 fleet is scaling, which is itself downstream of the Index data-collection pipeline described in Introducing Index: Building The World's Largest and Most Diverse Physical Dataset. So this doubles as a timestamped checkpoint on Figure’s hardware generation turnover.