Helix

Helix, unveiled by Figure AI in February 2025, is the company's in-house vision-language-action model built after ending its OpenAI partnership — the 'brain' behind the Figure 02 and 03 humanoids. It uses a System 1/System 2 architecture: a 7B vision-language model plans at 7–9 Hz while an 80M-parameter visuomotor policy executes continuous control at 200 Hz across 35 degrees of freedom, including individual fingers. Trained on roughly 500 hours of teleoperation data, a single set of weights handles picking, placing, drawer operation, and multi-robot collaboration — two robots were shown jointly tidying a kitchen from natural-language commands — all running on embedded onboard GPUs.

Limitations: proprietary and closed — no public weights, API, or independent benchmarks; all capability claims are Figure-reported. Whole-body loco-manipulation (Helix 02) remains an active research frontier rather than a solved product.