In-Depth with Figure AI: CEO Brett Adcock Dissects 60-Minute Logistics Demo, Touts Endurance Over Spectacle

SUNNYVALE, CA – "Jumped on to share the story behind our latest Helix upgrade," Figure AI CEO Brett Adcock announced, introducing a live YouTube roundtable, hosted by Herbert Ong, where he dissected the company's recent progress. Joined by a panel of investors and experts—including Chris Camillo (@chriscamillo), Scott Walter (@goingballistic5), Cern Basher (@CernBasher), and Amit Kukreja (@amitinvesting)—Adcock walked through a 60-minute, uninterrupted video of the Figure 02 humanoid performing a complex package sorting task. The discussion was framed as a deliberate statement on the company's shift in focus from brief clips to proving real-world endurance.

"We're here to sell humanlike work," Adcock stated during the stream. "That means running for hours every day and consistently hitting speed and performance targets. That’s what matters to our customers."

The demonstration, which the company had previously dubbed "the most boring video we've ever posted," shows the robot autonomously identifying packages on a conveyor, picking them up, and orienting them so a barcode faces down for an under-belt scanner—a task Adcock insists is impossible to solve with traditional heuristics.

Performance Metrics and Rapid Improvement

Throughout the discussion, Adcock provided KPIs for the task shown in the video, which was filmed the prior weekend. In that recording, the robot achieved an average package handling time of 4.05 seconds with a 95% success rate—already competitive with the 3-4 second rate of human workers.

However, Adcock revealed the pace of improvement is accelerating dramatically. In a follow-up announcement, he shared that just six days after the video was recorded, the same task was running at 3.54 seconds per package—a 13% speed-up. "The best part is the Helix neural network improvements are fed back and all robots using Helix benefit," Adcock noted, emphasizing the power of fleet-wide learning.

This rapid progress is happening while the robot's hardware is not yet being pushed to its limits. Adcock revealed the actuators on Figure 02 are currently operating at only 20-25% of their peak speed. "I don't see any reason we can't get to... beyond human-level performance in even this use case in the next 12 months in terms of speed," he projected.

The Helix Engine: Memory, Touch, and Data

The performance gains are powered by Figure's end-to-end neural network, Helix. Adcock detailed several recent architectural improvements that were critical to achieving this level of performance, aligning with technical details the company recently published.

Key upgrades to the Helix model include:

• Temporal Memory: The AI now incorporates several seconds of visual memory, allowing it to recall the state of the workspace even if it turns its head. "We used to forget," Adcock admitted.
• Force Feedback: The model now integrates force-sensing data, giving it a proxy for a sense of touch. This helps the robot perform more delicate manipulations, such as flattening a deformable polybag to ensure a successful scan—a subtle, human-like behavior learned from observation.
• More Data: The model shown was trained on approximately 60 hours of human demonstration data. Adcock emphasized that occasional hesitation or errors are "a data-starved problem" that will diminish as the data flywheel accelerates.

One of the most striking moments discussed was an emergent behavior where an engineer approached the robot and held out his hand. The robot, unprompted by a specific command, recognized the gesture and handed the package to him. "This was learned," Adcock confirmed. "There was some of this in training where the robot was handing things to humans... it'll do it to anybody [who] walks up."

Strategy: General Purpose, Co-Design, and the Race to Scale

The roundtable also delved into the strategic question of why a humanoid robot is necessary for a task that could, in theory, be handled by specialized automation. Adcock and the other panelists argued that the humanoid form factor provides the ultimate general-purpose solution for a world built by and for humans.

"You can make this argument on everything around the world," Adcock said. "We need maybe hundreds of millions or billions of robots that are all specialized... what we're trying to solve is the general-purpose version of this." He stressed that specialized systems are expensive, inflexible, and don't benefit from the "transfer learning" that will allow a single, advanced AI model to apply skills across countless domains.

He also pushed back against the notion that advanced AI can be a commodity simply placed on any hardware. "AI is not going to run well on shitty hardware," he stated bluntly, arguing that the AI and the physical robot must be co-designed. "You need to do the AI really well, you need the hardware right... I suspect there'll be like one or two groups outside of China that can really figure this out."

This vertical integration underpins Figure's core strategy: win by being the first to deploy a massive fleet. "The winner is... who can hit 100,000, who can hit a million units in the market as fast as possible," Adcock declared. "That's the winner. And that's a first-mover advantage, and speed means everything here." The data collected from such a fleet, he believes, will create an insurmountable lead.

With the next-generation Figure 03 already walking and designed for mass manufacturing, Figure AI is signaling it's not just building a capable robot, but a scalable system. That sentiment was echoed by panelist Scott Walter who, after a recent NDA-protected viewing of the new model, stated, "This is clearly the bot you want to scale... it's ready for prime time." Figure is betting that this combination of next-generation hardware and proven reliability on today's "dull, dirty, and dangerous" tasks is the most direct path to deploying millions of general-purpose humanoids into every aspect of the economy.

The full roundtable discussion, featuring further details on Figure's strategy and the panelists' reactions, can be viewed below.