Brett Adcock says Figure 4 will be the 'iPhone one moment' for humanoid robots — and data is the real bottleneck

Summary

Brett Adcock, CEO of Figure, argues that humanoid robotics has crossed a foundational technical threshold — and that data acquisition, not hardware or neural net architecture, is now the binding constraint.

Figure unveiled Helix 2 roughly three weeks before this conversation. The system runs fully end-to-end whole-body control using neural nets across more than 30 joints, processing camera input, robot state, and text or speech instructions onboard at roughly 200 cycles per second. Two onboard GPUs in the robot's torso handle inference entirely, with no external network required. The demo task was unloading and reloading a dishwasher — mundane by design. Adcock says the same stack could extend to laundry or other household tasks today if Figure had sufficient training data covering those task distributions.

The data bottleneck

Adcock is direct that the architecture is no longer the hard problem. The remaining gap between current capability and a general-purpose robot that can be deployed anywhere is data coverage. Figure plans to spend nine figures — implying at least $100 million — on data acquisition in 2026 alone. Teleoperation at scale isn't sufficient; Adcock says the data has to match the specific observation and action spaces of the Helix models, which rules out generic human demonstration data. Figure's answer is learning from humans at scale through what Adcock describes as the company's core model approach.

Separately, Figure has a major compute expansion going live April 1, adding to the hundreds of millions of dollars already deployed on training infrastructure. Training runs for Helix models are described as large and long.

Manufacturing cadence

Figure manufactures robots on its own campus. Figure 3 is currently coming off the line every three hours; Adcock expects that to reach one unit every 30 minutes within a few months.

China gap

On Chinese competitors, Adcock is skeptical. He argues that the absence of convincing neural-net-driven humanoid work from Chinese labs reflects real hardware limitations, not just a software lag. Unitree, for example, runs small reinforcement learning controllers on minimal onboard compute — insufficient to run Helix. Chinese robots also lack five-fingered hands, which Adcock treats as a prerequisite for human-like dexterity. He estimates Figure is at least a few years ahead of anything he's seen from China, while acknowledging that Chinese manufacturers have been valuable contributors to industrial robotics and consumer electronics supply chains more broadly. Figure designs and manufactures its robots internally and does not source designs from Chinese suppliers.

The stack milestone

For the first three-plus years of Figure's existence, Adcock says the company lacked a technical stack worth scaling. Early neural net work on tabletop manipulation was promising but narrow. The harder problem turned out to be leaving the tabletop — getting a full-body system walking, manipulating, and responding to prompts simultaneously in a coherent end-to-end loop. Helix 2 is Adcock's answer to that. He describes 2026 as the year Figure hits the accelerator on the back of a stack he's finally confident in scaling.