Frontiers · World models and embodied AI
You can describe the three layers of the 2026 humanoid stack — application companies, hardware platforms, and foundation models — and place the major players in each layer.
"Humanoid robotics" got crowded fast in 2026. The right way to read the field is as three layers, not one race.
Application-and-deployment companies build humanoid robots and put them into customer environments. Figure AI deployed Figure 02 in BMW Spartanburg starting 2024 and reported 1,250 operational hours and 90,000+ sheet-metal load operations by mid-2026; the next-gen Helix 02 vision-language-action model runs on board with two NVIDIA RTX GPU modules. 1X Technologies began NEO consumer-humanoid pilots through 2025-2026. Tesla Optimus targets 50K-100K units in 2026 and 10M/year at Giga Texas by 2027 (the second figure is a Musk projection, not a milestone). Agility Digit is in distribution-center pilots with GXO. Boston Dynamics Atlas (electric) is in commercial pilots with Hyundai.
Hardware platforms sit underneath. Unitree shipped ~5,500 units in 2025 and is targeting 10K-20K in 2026; the G1, H1, and R1 humanoid lines plus the Go2 quadruped form the most-shipped humanoid platform by unit count. The hardware-platform layer is where China-based manufacturers dominate.
Foundation models sit underneath both. NVIDIA GR00T N1.6 shipped January 2026 as an open foundation model for humanoid manipulation; Gemini Robotics-ER 1.6 ships through Apptronik integrations and CES 2026 Boston Dynamics partnerships; Physical Intelligence's π-0.7 model shows first signs of compositional generalization across embodiments; Skild AI's "Skild Brain" pitches general robot foundation models.
Chapter contains 3 lessons.