Figure AI

Publishes technical system details and real-world evaluations for generalist humanoid control.

No.Paper · what it does and showsIndustry contribution

Introducing Helix 02: Full-Body Autonomy2026 ↗

Extends Helix to the whole body with a learned System 0 controller trained on over 1,000 hours of human motion, completing a continuous 4-minute, 61-action dishwasher task with walking, manipulation and balance and no resets.

Replaced 109,504 lines of hand-engineered whole-body C++ control with a neural prior and added tactile and palm-camera dexterity such as single-pill extraction and 5 ml syringe dosing.

Helix: A Vision-Language-Action Model for Generalist Humanoid Control2025 ↗

A single VLA that controls the humanoid’s upper body from pixels using a dual-rate system: fast visuomotor reflexes at high frequency plus a slower reasoning VLM, all running on onboard hardware.

One of the first humanoid VLAs running onboard in commercial pilots, showing generalist humanoid control can live on consumer-grade compute.

Scaling Helix: A New State of the Art in Humanoid Logistics2025 ↗

Adds vision memory, state history and force feedback and scales demonstrations from 10 to 60 hours, handling poly bags and envelopes at about 4.05 s per package with roughly 95% barcode-scan success on a live conveyor.

Benchmarked learned humanoid manipulation against human speed and reliability inside a real distribution centre, the evidence logistics buyers require before ordering fleets.

Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot Transfer2025 ↗

Builds an internet-scale humanoid pretraining dataset through a Brookfield partnership covering over 100,000 residential units, and shows zero-shot transfer from 100% egocentric human video to language-driven robot navigation with no robot demonstrations.

Establishes large human-video data partnerships across homes and commercial real estate as the strategy for humanoid pretraining.