EXCLUSIVE: Brett Adcock on Figure’s 8-Hour Autonomous Shift

In this talk

1. 1:00

   Autonomous package sorting

   Adcock says Figure has run this setup for months using Helix 2, which maps camera input to whole-body movement. Robots handle packages at roughly three seconds each and about 90% success, while coordinating charging and maintenance handoffs to keep the conveyor moving.

2. 5:42

   Training data and shared learning

   Adcock describes Helix 2 as an in-house vision-language-action policy trained on collected data, with inference running onboard. Figure uses teleoperation and human video among its data sources. He says adding varied non-fridge examples raised fridge-task success from about 55–60% to around 90%, supporting a shared model trained across tasks.

3. 18:30

   Figure 4, hands, and supply chain

   Adcock says Figure had completed a major design review for Figure 4 and calls it a substantial redesign built around Helix 3 and data collection. He says the new hand is designed to match human range of motion and that human-like hand mechanics could help robots learn from human demonstrations. Figure designs most robot components internally and has moved most of its supply chain out of China.

4. 25:08

   Constraints and general-purpose robotics

   Adcock identifies data as Figure’s main current constraint, followed by energy, compute, and manufacturing. He says the company has brought a new NVIDIA B200 cluster online for large-scale model training. Figure aims to make robots handle unfamiliar tasks and settings; Adcock says a demonstration could come as soon as this year, while estimating broader progress may take one to three years.

5. 29:03

   Home use and human environments

   Adcock says Figure wants to deliver a dependable product rather than ship robots before they work well. He describes a home test in which a robot loaded laundry, but says robots remain far from handling everyday household tasks reliably. He argues humanoid form fits spaces and tools built for people, and says lighting changes can reduce policy performance unless training data covers those conditions.