Originally published on The Daily Flare.
Humanoid robots in 2026 are no longer science fiction, but not coworkers yet either. This tracker covers what they're actually doing — factories, warehouses, homes — and the honest gaps between demos and everyday work.
A decade ago, a humanoid robot was a research project with a kill switch. Now machines shaped like people are clocking shifts on factory lines, in warehouses, and in living rooms on a subscription. This tracker maps the state of humanoid robots 2026 — what eight major programs are actually doing. Targets are labeled as targets; unverified company claims are flagged.
Deployment milestones for humanoid robots 2026
One rule governs every entry below: real deployment is narrow — one task, one place, simple grippers.
| Company | Robot | Where it works | Headline result |
|---|---|---|---|
| Tesla | Optimus | Tesla's own factories | Internal only; R&D phase (reported) |
| Figure AI | Figure 02 / 03 | BMW plants (US, Germany) | 10-month pilot; 30,000+ cars supported |
| Agility Robotics | Digit | GXO warehouse (GA), plus pilots | First commercial humanoid RaaS contract |
| Apptronik | Apollo | Mercedes-Benz plants | Logistics pilots; product due 2027+ (target) |
| Boston Dynamics | Electric Atlas | Hyundai Georgia Metaplant | Parts-sequencing pilot (plan) |
| 1X Technologies | NEO | Preorder stage | First US deliveries scheduled 2026 (company plan) |
| Sanctuary AI | Phoenix | Pilot scale | Magna alliance; latest model dropped legs for wheels |
| Unitree | G1 / H1 / H2 | Research labs, mostly | 5,500+ shipped 2025 (self-reported); 70%+ to research |
Tesla Optimus illustrates the announcement-to-deployment gap best. It works only inside Tesla's own facilities. Musk targeted 5,000 units for internal deployment in 2025; actual output was a few hundred — under 10% of the goal, per reporting by The Information. Musk said the ~1,000–1,200 deployed units were "still very much in the R&D phase" (Q4 2025 call), and admitted earlier public demos had been preprogrammed or remotely controlled (Q2 2026). Reported figures (September 2026) put output at several hundred per week by August, targeting above 1,000 per week by end of 2026. Musk's price target is $20,000–$30,000 per unit, against an estimated current build cost of $50,000–$100,000-plus.
Figure AI has the best-documented production deployment of any humanoid. In a pilot announced by BMW Group in February 2026, Figure 02 ran a ten-month trial at BMW's Spartanburg plant in 2025 — the first humanoid deployment at a BMW factory. Working ten-hour shifts, the robots placed sheet-metal parts within a 5 mm loading tolerance, supporting production of more than 30,000 BMW X3s and moving 90,000+ components over about 1,250 operating hours. Because the customer co-signs the numbers, this is the industry's most credible deployment data. BMW has since started a European pilot at Plant Leipzig, and in summer 2026 the newer Figure 03 moved into Hall 52 at Spartanburg for logistics sequencing. Figure's bolder claims — a robot every ~90 minutes off its BotQ line, 50,000 units in 2026 — are company-reported ambition, not audited output.
Agility Robotics' Digit holds the industry's first formal commercial deployment of a humanoid. After a late-2023 proof of concept, GXO signed a multi-year Robots-as-a-Service agreement in June 2024 — also the first humanoid RaaS contract — putting Digit to work at a GXO-run Spanx facility in Flowery Branch, Georgia, picking totes from autonomous mobile robots onto conveyors. Unit counts are undisclosed ("small fleet"). Reported follow-on agreements include Schaeffler (November 2024), Mercado Libre in San Antonio (December 2025), and Toyota Motor Manufacturing Canada — seven Digits at its Woodstock, Ontario RAV4 plant (February 2026) after a year-long pilot. Digit's specs: ~1.75 m, 65 kg, 16 kg payload, three to five hours per charge with auto-docking. Agility deliberately uses logistics-specific grippers instead of dexterous hands — the narrowness that makes its deployment work.
Apptronik's Apollo has been running a long pilot with Mercedes-Benz under a commercial agreement announced in March 2024: logistics work — assembly-kit delivery, kitted-tote delivery, component inspection — at Berlin-Marienfelde and Kecskemét, Hungary. Mercedes-Benz invested a "low double-digit million" euro amount (reported via German press); operators trained Apollo via teleoperation and AR. Apptronik raised $520M in February 2026 — backed by Google and Mercedes-Benz at a reported ~$5B valuation. Apollo 2 was unveiled publicly on June 30, 2026, but CEO Jeff Cardenas told Reuters the company had already built hundreds of units that had "worked for more than a year" in pilots before the reveal; its ~90,000-square-foot Austin "Robot Park" is a data-collection and training facility. The real commercial product is Apollo 3: pilots continue through 2026, production versions expected "in 2027 and beyond" — a company target. Apollo's deployments still run under "segregated safety" — light curtains and external sensors that halt the robot when a human enters its workspace.
Unitree ships more humanoid robots than anyone: 5,500+ in 2025, per its March 2026 Shanghai STAR Market IPO filing, with revenue up ~335% to ¥1.71B (~$250M). The filing claims a 32.4% share of global humanoid unit shipments — disputed by independent analyst Omdia, which credits AgiBot with 39%. The decisive caveat: an estimated 70%+ of 2025 humanoid revenue came from research institutions — developer machines, not factory workers. Prices remain the industry's floor: the G1 at roughly $13,500–$16,000, the H2 from ¥199,000 (under $30K), the R1 from ¥39,999 (~$4,900). CEO Wang Xingxing was candid at the 2026 World Robot Conference: Unitree had held back from large-scale factory deployment because throughput still lags human workers and every new task requires retraining. One regulatory wrinkle: the FCC barred Unitree from North America in 2026 (reported), tilting its geography toward domestic China.
1X's NEO is the only major program aiming straight at consumers. The early-2025 plan to pilot NEO in "a few hundred" real homes by end of 2025 never received public confirmation of completion. Preorders opened at the October 28, 2025 launch: $20,000 outright or $499/month (six-month minimum), first US deliveries scheduled for 2026. 1X claims it sold out its entire first-year capacity — 10,000 units — in five days; company-claimed and unverified. NEO stands 5'6" (168 cm), runs three to four hours active plus eight on standby, and uses tendon-driven actuators marketed as low-force for home safety. CEO Bernt Børnich is candid: complex tasks are handled by remote 1X operators wearing VR headsets — teleoperation, not full autonomy. "Not everything will work… Some will be autonomous, some will be tele-op. And sometimes it will fail." The privacy question — cameras throughout the house — is structural; 1X answers it with "safety guardrails," an unaudited marketing assurance.
Sanctuary AI's Phoenix remains pilot-scale. A strategic alliance with Magna (April 2024) set a roadmap for production batches of Phoenix robots in real manufacturing environments — but no unit counts or results have been disclosed. In 2023, Phoenix was documented performing about 110 tasks under real conditions at a Canadian Tire store — company-documented, with no independent verification. Phoenix Gen 7 (July 2026) claims sub-24-hour task learning — self-reported. Most telling is Gen 8, which dropped bipedal legs for a wheeled base after they proved too frail.
Boston Dynamics' electric Atlas is the captive-deployment model: Hyundai owns 88% of the company, and Atlas's route to work runs through Hyundai's plants. A pilot began in fall 2025 at Hyundai's Georgia Metaplant in Savannah — parts-sequencing field tests with a prototype fleet, continuing into early 2026. The product version unveiled in October 2025 is 1.9 m tall, lifts up to 30 kg repeatedly, runs about four hours, and self-swaps its battery in under three minutes. The sequence ahead: 2026 production goes to a training center (RMAC) and to Google DeepMind — a "data factory" phase; early 2027 brings the first external customers (announced intent, none named); 2028 brings production-line deployment at the Metaplant, committed by Hyundai but "subject to continued technology validation." Hyundai's fleet plan speaks of 25,000+ Atlas units and 30,000-per-year factory capacity — plans, not output.
The honest gaps between demos and useful work
The deployment record is real progress wrapped in unaudited claims. Here is what separates what works from what marketing sells.
Battery life and uptime. The practical ceiling is three to five hours per charge: Digit at three to five, Apollo about four on swappable packs, Atlas about four with its sub-three-minute self-swap, NEO three to four active. No current humanoid works a full shift untethered.
Dexterity versus gripper simplicity. The two robots genuinely deployed at scale — Digit and Figure 02 — succeeded precisely by avoiding general dexterity: tote carrying and sheet-metal placement with simple grippers. Fine manipulation — Tesla's 22-degrees-of-freedom Gen 3 hand, Sanctuary's hydraulic hands — remains lab-demo grade, with no factory-hour data behind it.
Autonomy versus teleoperation. The industry's quiet secret is how much "autonomous" footage is human-driven. Musk admitted Optimus's earlier public demos were preprogrammed or remotely controlled; 1X's CEO openly sells NEO as part autonomous, part tele-op; Apptronik trained Apollo largely through teleoperation plus AR. True autonomy exists only in narrow, pre-mapped tasks.
Safety certification. Nobody has a certified-collaborative humanoid on a live mixed line. Apollo still runs segregated safety; Figure 03's soft coverings are a design gesture, not certification. Regulators and insurers, not AI, set the deployment pace.
Cost versus wage parity. Figure prices its roadmap to "warehouse wage parity" through a Robots-as-a-Service model; Digit sells as-a-service for the same reason. Optimus's $20–30K target assumes scaling Tesla has not demonstrated. Unitree's sub-$16K G1 looks cheap — but it is a research platform, not a worker.
Throughput. Robot throughput still lags human workers, and each new task needs retraining — Unitree's CEO said so. BMW's pilot worked because the task was one repetitive placement.
Verification. No independent body audits "hours worked" or "tasks completed." Livestreams and company metrics are company-reported. Where a customer co-signs — BMW Group, GXO — the claim is materially stronger.
| Gap | State of play | Closest so far |
|---|---|---|
| Uptime | 3–5h ceiling; no full shift untethered | Atlas self-swap (<3 min), Apollo hot-swap |
| Dexterity | Simple grippers win; fine hands are lab demos | Figure 02, Digit (tote handling) |
| Autonomy | Narrow tasks only; much footage is teleop | Figure 02 at BMW (10-month pilot) |
| Safety | No certified-collaborative humanoid on mixed lines | Apollo segregated safety (honest, limited) |
| Cost | RaaS toward wage parity; purchase prices still high | Digit (RaaS), Figure roadmap |
| Throughput | Lags human workers; retraining per new task | Nobody — Unitree's CEO said so |
| Verification | No independent audits; customer co-signs count | BMW Group, GXO data |
What's next on the road to everyday work
The stated targets for 2026–2027 tell you where the money is going, if not what will actually arrive. Tesla's reported internal target is an automated line above 1,000 Optimus units per week by end of 2026, with public purchase floated for late 2027. Figure's company target is 50,000 units in 2026 from BotQ. 1X plans first US consumer deliveries of NEO in 2026. Apptronik's Apollo 3 production versions are expected "in 2027 and beyond." Boston Dynamics aims at its first external Atlas customers in early 2027 and production-line deployment at the Georgia Metaplant in 2028. Unitree targets 20,000 humanoid deliveries in 2026.
None of these are commitments. Nearly every target here has already been revised: Optimus's 5,000-unit 2025 target was missed at under 10%; Apollo's commercialization date slid from "2025 general availability"-era statements to Apollo 3 in "2027 and beyond." Treat stated targets as direction of travel.
The measured answer: in 2015, a humanoid was a lab demo. In 2026, humanoids are doing paid work in narrow lanes — one repetitive task each, under human supervision and safety fencing. Useful everyday work — a full shift untethered, autonomous handling of varied tasks, certified to work beside people — exists nowhere yet. But for the first time, the pipeline from pilot to production line is visible — and the companies being honest about what they haven't solved are making the most progress.
Top comments (0)