The Humanoid Race:
Robotics Q2 2026

Figure AI "units in field" reflects F.02 retired post-BMW pilot; F.03 entering commercial deployment Q2–Q3 2026. Tesla's headcount reflects the full AI/robotics org; Optimus-specific team estimated at 500–1,000. Boston Dynamics 2026 Atlas deployments are fully committed — no new customer capacity until 2027. Agility's 100+ commercial units is the highest confirmed deployment count among US startups.

This is the most analytically underexamined event in the humanoid robotics space. The decision to terminate the OpenAI partnership and build Helix in-house sets a precedent the entire intelligence layer debate turns on.

January 2024: Figure AI announces a formal partnership with OpenAI. The deal was positioned as transformative — OpenAI would build specialized AI models for Figure's robots, enabling natural language interaction and high-level task planning. Figure raised $675M in the same month (Series B) with OpenAI's startup fund as a direct investor, alongside Microsoft, NVIDIA, Jeff Bezos, Amazon, and Intel Capital. Valuation: $2.6B.

Early 2025: The partnership quietly ends. Figure announces Helix, its proprietary Vision-Language-Action model. CEO Brett Adcock explains the shift bluntly: "Large language models had become a smaller problem compared to those allowing for high rate robot control."

In other words: the hard problem in robotics isn't language understanding. It's getting a robot to move a physical object at industrial precision with millisecond response times, thousands of times per hour, without dropping it or damaging equipment. GPT-4o doesn't help you there.

Helix is a dual (now triple) system VLA architecture running fully onboard:

The key innovation: the same model handles tasks across different contexts with no task-specific retraining. New capabilities are added through data — not new algorithms. Helix 02's ability to sequence 60+ actions in a 4-minute kitchen task, autonomously, without resets, would not have been achievable through an API-layer integration with OpenAI.

The Figure pivot signals that foundation model providers don't own the robot layer — robot companies do. OpenAI can't put its model into a proprietary torque-density actuator. Figure can. This is the same logic that drove Tesla to kill Mobileye in 2016 and build its own vision system.

The $20–30K per-unit price target is cited by every Western humanoid company. The reality of current cost structures makes it clear that only one company has a credible path there before 2028.

Who gets to sub-$30K BOM by 2027: Tesla, almost certainly. 1X is already there in hardware cost but trades off autonomy. Figure is on track if BotQ executes. Everyone else is 2028+.

This is the most credible industrial pilot result released by any Western humanoid company. 11 months, quantified throughput, disclosed failure modes, real production contribution. The benchmark everyone else is measured against. BMW has not publicly confirmed expanded F.03 deployment, though Figure announced a second major commercial partner (undisclosed, confidential) alongside BMW.

Apollo deployed at Mercedes manufacturing facility. Announced March 2024 as the first publicly disclosed commercial deployment by Apptronik. Performance data: minimal public disclosure. CEO Jeff Cardenas told TechCrunch (early 2025): "we have yet to move beyond the pilot stage" with Mercedes or any partner.

Strongest signal: Mercedes-Benz led the $53M extension round of Apptronik's Series A (March 2025). If the pilot had failed, they wouldn't have written a check. Assessment: Real deployment, limited public metrics, strategic investor commitment as proxy for confidence.

Agility is the only humanoid company with genuine, documented commercial deployments generating recurring revenue.

Assessment: Industry-leading commercial validation by deployment count and contract structure. The RaaS model is strategically brilliant — recurring revenue, lower adoption barriers, continuous training data. Downside: Digit is purpose-built for logistics tote-handling. Won't compete with Figure or Tesla in general-purpose manufacturing.

~800–1,000 Optimus Gen 2 units deployed internally at Tesla manufacturing facilities. Tasks: battery cell sorting, parts handling, quality inspection. As of early 2026, units are "primarily engaged in learning and data collection rather than productive manufacturing tasks." Tesla explicitly frames the factory deployment as a proving ground: "We look at the factory as a lab."

Assessment: The scale bet is real — Fremont factory converting to Optimus production, Q2 2026. But current economic contribution of Optimus inside Tesla is near-zero. This is still a capital investment, not a deployed product. The narrative is two years ahead of the reality.

Limited deployments in "a few hundred to a few thousand" homes by end 2025 (CEO Bernt Børnich, GTC 2025). Journalist Joanna Stern's WSJ coverage revealed that "most tasks were teleoperated by a human with a virtual reality headset." The honest framing: this is a training data collection program disguised as a consumer product launch.

Assessment: Technically behind Figure and Apptronik on autonomy. Strategically ahead on consumer data diversity. Privacy concerns are real — NEO streams video and audio from inside homes.

Atlas entered training phase at Hyundai's RMAC in Georgia. Production-ready Atlas unveiled at CES 2026 — all 2026 deployments already committed, zero spare capacity for new customers until 2027. Full commercial deployment at Hyundai EV facility (Ioniq 5/9) not until 2028. Training new tasks reportedly takes one to two days with the NVIDIA-powered AI brain.

Assessment: The most credible hardware platform — 30+ years of engineering, 56 DOF, 50 kg lift, industrial-grade tolerance. The limitation is pace. Atlas is methodical excellence at automotive timeline, not nimble startup execution.

The humanoid robotics race will ultimately be decided not by which company has the best hardware this quarter, but by who controls the best training data in 2027–28. VLA models require continuous, diverse, high-quality teleop and autonomous interaction data. This is the equivalent of the LLM pre-training corpus question — except the data is physical, not textual.

Three named, time-bound, falsifiable predictions for 2026.