Figure All-Hands Meeting Reveals Roadmap: Figure 04 Enters Component Manufacturing, Betting on Large-Scale Pretraining and $3.5 Billion in Compute

Main body of the article

Recently, the American humanoid robot company Figure publicly released a one-hour internal all-hands meeting video. At this meeting, there was no flashy new product showcase and no grand long-term blueprint; instead, it depicted an increasingly clear strategic path—as hardware enters a period of iterative convergence, how Figure will implement the engineering route of "large-scale pretraining + massive compute" to make robots gradually "smarter." The full video was posted on the company's official YouTube channel on October 22 local time.

Specifically, Figure AI founder and CEO Brett Adcock reviewed the company's current R&D foundation, disclosed the latest progress of the fourth-generation humanoid robot Figure 04 and the Index global human behavior data collection project, details of the $3.5 billion compute partnership, and the internal testing of the new-generation Helix model.

Figure 04 enters the parts manufacturing stage; the global data collection project becomes a moat

On the hardware side, the latest-generation robot Figure 04 has completed design lock and its parts have entered the delivery and validation cycle. Compared with Figure 03, which has already deployed logistics sorting tasks on BMW production lines and achieved continuous long-duration autonomous operation, Figure 04 is not simply optimizing battery life and payload; rather, it recalibrates the actuation units around the manipulation requirements of the Helix model, and a new generation of actuators, batteries, and dexterous hands has simultaneously entered the reliability testing stage. Among these, the dexterous hand is one of the areas into which the R&D team has put the most effort.

Adcock called Figure 04 the largest improvement of any generation, and stated clearly at the meeting that the mass-production deployment of Figure 03 has already completed hardware feasibility validation, while the goal of Figure 04 is to build a standardized hardware foundation suitable for general household and complex mixed industrial scenarios, supporting the model in stably executing long-sequence tasks in more unfamiliar environments.

The foundation supporting model iteration is a global data collection project called Index, which is also a unique path that distinguishes Figure from most peers. Index relies on a collection app open to users worldwide, recruiting ordinary people to record daily operation behaviors in scenarios such as homes, offices, and warehouses, rather than relying on manual collection in laboratories or teleoperation data recording.

As of this meeting, the project has accumulated 37.5 million human operation videos, covering more than a hundred countries, and the data collection rate has increased to about 81 minutes of video per second, equivalent to adding about 116,600 hours of video content per day. Massive long-tail scenario data is key to improving the zero-shot generalization capability of the Helix model. Previously, Helix 2.5, relying on Index pretraining, increased its task completion rate from 9% to 56% in never-before-seen home environments.

However, Adcock also admitted that this route is costly, with weekly data-related investment reaching $8.1 million. In his judgment, over the next 12 months, the company's investment in data collection may exceed all the funding most robot companies have obtained to date.

$3.5 billion compute partnership + self-developed large models "escort" product iteration

Compute is the foundation for further iteration of its robots.

Adcock's logic is that there are many types of real-world tasks, and it is impossible to collect data and post-train models separately for each job. Therefore, it is necessary to first use large-scale human behavior data for pretraining, and then use a small amount of task data to adapt to specific skills. To digest exponentially growing training data, Figure has finalized a compute partnership with AI compute cloud provider Nscale with an initial scale of $3.5 billion, planning to deploy up to 100,000 GPUs based on Nvidia's Vera Rubin platform. The compute resources are planned to come online in the second half of 2027 at the Bastrop, Texas site, and the subsequent partnership scale is expected to expand to $6 billion.

At the end of the speech, Adcock revealed that the team had trained a new model using the largest-scale pretraining data and GPU resources to date, and completed testing a few weeks ago in an internal lab codenamed "Black Box."

He did not give any evaluation details, saying only that he had seen robot performance never seen before, calling it one of the most important moments since he began robotics research, and describing his early feeling as "Baby AGI." The team is training a larger-scale version, with new progress expected within weeks, and plans to carry out even larger-scale training in the next six months.

(Source: Science and Technology Innovation Board Daily) Middle of the main body: embedded advertisement