Written by Marcus Chen
Figure has released Helix 2.5, an AI model for its Figure 03 humanoid that lifted whole-task success in 30 unfamiliar homes to 56 percent, against 8 percent for an otherwise identical system without the new pretraining. The result targets robotics’ long-standing “holy grail”: carrying learned skills into places a robot has never seen.
A humanoid robot that works in one factory is a demonstration. One that works in a factory it has never entered is a product. Figure, the San Jose company founded in 2022 by Brett Adcock, has now published evidence that it can close part of that gap.
Helix 2.5, an upgrade to the AI system controlling Figure 03, was evaluated in 30 unfamiliar homes around the San Francisco Bay Area. Robots tidied living rooms, made beds and folded towels without any data collected in those homes, encountering unfamiliar furniture, objects and layouts while coordinating vision, walking, balance and two-handed manipulation. Under Figure’s strict criterion — completing an entire task — Helix 2.5 achieved a 56 percent zero-shot success rate. An otherwise identical system without the new pretraining managed 8 percent. The model also needed roughly half the task-specific data of the earlier Helix 02 generation.
Figure attributes the gain largely to Index, its human-activity dataset, which the company describes as the largest and most diverse physical dataset for robot training. More than 90,000 people contribute each week, adding about 35 minutes of human experience every second. Performance improved consistently as Index data increased, hinting that robot learning may follow the scaling behaviour that made large language models work.
The hardware is built for the same ambition. Figure 03 stands 173 cm, weighs 61 kg and carries up to 20 kg, runs about five hours and walks at up to 1.2 m/s. Its cameras offer twice the frame rate and a 60 percent wider field of view per camera than Figure 02, palm cameras support close manipulation, and fingertip sensors resolve forces as small as three grams. Figure has committed an initial $3.5 billion to AI infrastructure, potentially exceeding $6 billion, and a partnership with Nscale is to supply as many as 100,000 NVIDIA GPUs from the second half of 2027.
Generalization is arriving alongside a harder problem: fleet-scale deployment. On 9 October, Hexagon Robotics described putting at least 1,000 of its Aeon humanoids into German manufacturer Schaeffler’s global factory network by 2032, an arrangement its president, Arnaud Robert, argues must be planned as a fleet rather than as 1,000 copies of one robot. Hexagon’s training loop feeds customer-specific, sensor-rich data back into simulation — a “sim-to-real-to-sim” cycle — while insisting that factories cannot rely on teleoperation. Its Aeon is an unusual wheeled biped, chosen because wheels are faster on flat floors and legs provide access to everything else.
The commercial backdrop is shifting quickly. Tesla has signalled a target of more than 1,000 Optimus units per week by the end of 2026, and IDC counted nearly 25,000 humanoid shipments in the first half of 2026, up 432 percent year over year. Venture funding for humanoid startups reached $11.0 billion this year, more than double 2025’s full-year record.
The caveats are real. A 56 percent success rate on complete tasks is not reliability, and Figure’s 30-home study is small and self-reported. Sceptics note that demonstrations scale faster than reliability does. But the gap between 56 percent and 8 percent is the interesting number: it suggests generalization is becoming an engineering property that improves with data, rather than a permanent barrier. If that holds, the humanoid industry’s next competition will be over datasets, not joints.
