Perception for chaos

Robots in warehouses do not just need to see boxes; they need to keep track of the same box as it slips behind other packages and reappears a second later. Figure said its latest perception work is aimed at that messier, real warehouse problem, not a clean lab setup. (figure.ai) Figure’s logistics system, called Helix, is built for package manipulation and triaging on moving conveyors, where parcels vary in size, shape, weight, and rigidity. In its February 26, 2025 post, the company said the robot must track a dynamic flow of numerous packages while keeping labels oriented for scanning at high throughput. (figure.ai) The company’s June 7, 2025 update tied better results to a “vision memory” module and a history of past states, which let the system hold onto context across time instead of treating every video frame as a fresh scene. Figure said that stateful setup improved robustness to interruptions while handling deformable poly bags, padded envelopes, and rigid boxes. (figure.ai) That is the basic perception problem in plain terms: if a package is partly hidden by clutter or another package, the robot still has to infer that it is looking at the same object. Figure’s posts describe that as temporal memory and stateful perception, the software equivalent of not losing your place when something passes behind a shelf. (figure.ai) Figure said those changes showed up in warehouse-style metrics, not just demos. In the June 2025 report, it said execution speed improved to 4.05 seconds per package from about 5.0 seconds, while label orientation for barcode scanning rose to about 95% from about 70%. (figure.ai) The company also said training data grew from 10 hours to 60 hours in that update, alongside force feedback and model changes. In an earlier February 2025 post, Figure said as little as 8 hours of well-curated demonstration data could produce a flexible policy for this logistics task. (figure.ai 1) (figure.ai 2) Figure has framed warehouses as a proving ground for neural networks because the scene changes every moment and simulation is hard. Its June 2025 post said small-package logistics is “perfectly suited” for learning-based systems because packages crumple, flex, and arrive in constantly changing arrangements. (figure.ai) The broader push is commercial, not academic. Figure said on February 20, 2025 that Helix runs onboard low-power embedded graphics processors for deployment, and on February 26, 2025 it introduced logistics as a real-world application for its humanoid robots. (figure.ai 1) (figure.ai 2) Figure’s public materials still leave gaps, including how the latest warehouse perception results compare with rivals under the same test conditions. But the company’s own posts make the target clear: a robot that can keep object identity intact through clutter, occlusion, and conveyor-belt chaos long enough to keep sorting. (figure.ai 1) (figure.ai 2)