Three years ago, humanoid robots were lab curiosities — slow, brittle, hand-programmed for single tasks in fixtured environments.
Today: 13,000 shipped in 2025. One model walked 106 km to set a Guinness World Record. Another ran 10-hour production shifts at BMW for eleven months.
The transition from demo to deployment happened faster than almost anyone predicted. Here’s what the field actually looks like now.
The Numbers That Tell the Story
- 13,000 humanoid robots shipped globally in 2025 (entire industry)
- 5,168 shipped by Agibot alone — #1 globally, in a single year
- 90,000 parts loaded by Figure 02 at BMW Spartanburg over 11 months
- 98% task success rate for Agility Digit at Amazon after 18 months
- $4,900 — the cheapest humanoid robot ever, Unitree R1 Air (April 2026)
- 140+ Chinese humanoid manufacturers as of end-2025
- $39B — Figure AI’s valuation in September 2025
The scale numbers are early but real. The cost numbers are where the story gets interesting.
What They Can Actually Do
The capabilities question has two parts: physical and cognitive.
Physical
Dexterity is the bottleneck that everyone was watching. Human hands have ~27 degrees of freedom. The leading robots now reach 16–22 DOF per hand:
| Robot | Hand DOF |
|---|---|
| Tesla Optimus Gen 3 | 22 DOF per hand |
| 1X Neo | 22 DOF (tendon-drive) |
| Figure 02 / 03 | 16 DOF, grip adaptation at 200 Hz |
| Fourier GR-2 | 12 DOF + 6 tactile sensor arrays |
That last one matters: DOF count is no longer the bottleneck. Force sensing and material discrimination are the next frontier.
Payload ranges from 5 kg (Agibot G2 single arm) to 25 kg (Apptronik Apollo — highest in class). Most warehouse-class robots sit in the 16–20 kg range — sufficient for tote handling, parts loading, box movement.
Speed has improved: Unitree H1 holds the humanoid speed record at 3.3 m/s. Most industrial units run 1.2–1.5 m/s — not fast, but fast enough for most logistics tasks.
Cognitive: The Foundation Model Shift
This is what changed everything. Rule-based robots had lookup tables. If the part is in position X, move the arm to position Y. Every variant required engineering.
Vision-Language-Action (VLA) models replaced that. Input: images + natural language. Output: motor commands. The robot reasons about what to do next rather than pattern-matching against a script.
Key models in production:
π0 / π0.5 (Physical Intelligence) — trained on 10,000+ hours across 7 robot types and 68 tasks. Open-sourced in February 2025. π0.5 adds open-world generalization.
Helix (Figure AI) — Figure’s proprietary VLA, replacing their OpenAI collaboration. Zero-shot generalization to thousands of novel objects not seen in training.
GR00T N1 (NVIDIA) — open foundation model for humanoid robots. 2B parameters.
What this unlocks: robots can now follow natural language instructions, generalize to new objects and environments, chain multi-step tasks (“clear the table”), and recover from failures using learned heuristics instead of stopping.
What it doesn’t unlock yet: sustained high-speed repetitive electronics assembly, true unstructured outdoor environments.
The Cost Story
This is where the competitive dynamics get strategic.
Current price range spans three orders of magnitude:
| Robot | Price |
|---|---|
| Unitree R1 Air | $4,900 |
| Unitree G1 (base) | $13,500–$16,000 |
| Tesla Optimus (projected at scale) | $20,000–$30,000 |
| Apptronik Apollo (est.) | $50,000–$100,000 |
| Figure 02 | $100,000+ |
| Agility Digit (fleet) | ~$250,000 |
The critical number: Chinese BOM (bill of materials) runs ~$46K for a full-size humanoid. Western supply chain: ~$131K. Nearly 3x difference. That’s why Unitree can sell at $5,900 and Western manufacturers cannot.
Tesla reduced the cost of a single actuator component (planetary roller screw) from ~$3,000 to ~$800 — 75% reduction on one part. This is the playbook being applied across all actuators. The trajectory toward sub-$20K is real; the question is how fast.
RaaS Economics
Outright purchase is the current dominant model, but Robot-as-a-Service is growing fast. At 3-shift operation, industrial RaaS at $3,500–$5,000/month works out to $5–7/hour effective cost — against $25–35/hour for US manufacturing and logistics labor. That’s where the ROI math already closes.
1X Neo at $499/month is a different play: quality-of-life for consumers, not economic ROI. That’s the pricing at which consumer robotics becomes a real conversation.
The Competitive Landscape
US Players
Tesla Optimus Gen 3 (production started January 21, 2026): 22 DOF hands, 50 total actuators, 8 cameras. 1,000+ units deployed in Tesla factories. Musk’s Q4 statement: they’re not doing “useful work” yet — all in learning/data collection mode. Production target: 50,000–100,000 units in 2026, 1M/year by 2027.
Figure AI: Figure 02 ran 10-hour shifts at BMW Spartanburg for 11 months — 90,000+ parts loaded, 30,000+ vehicles contributed to. Valuation went from $2.6B to $39B in 18 months. Figure 03 announced October 2025; BotQ facility tooled for 12,000 units/year.
Agility Robotics (majority owned by Amazon): Digit at Amazon Sumner — 98% task success rate, effective cost $10–12/hour, 100,000+ totes moved. The most documented sustained industrial deployment in the field.
1X Technologies (Norwegian, OpenAI-backed): Neo at $20K or $499/month, 22 DOF tendon-drive hands. Industrial pivot: deal for up to 10,000 units in EQT portfolio company facilities.
Boston Dynamics Atlas: 6’2”, 360-degree joint rotation. All 2026 deployments committed to Hyundai and Google DeepMind. Not yet at production-scale autonomous manipulation.
Physical Intelligence: Not a robot — a foundation model (π0) that runs on diverse hardware. Funded at $5.6B. Business model: license the model to OEMs.
Chinese Players
Unitree Robotics: Sets the global price floor. R1 Air at $4,900 is the most disruptive pricing event in robotics history. G1 ($13,500+), H1 holds the speed record at 3.3 m/s.
Agibot: Shipped 5,168 robots in 2025 (#1 globally). Produced its 10,000th humanoid on March 30, 2026 — scaled from 5,000 to 10,000 in three months. Walked 106 km for a Guinness record. US market entry announced.
Fourier Intelligence: GR-2 has 53 total DOF, 6 tactile sensor arrays sensing force, shape, and material. Founded in rehabilitation robotics — their sensing work is ahead of most Western competitors.
UBTECH: Walker X/S. 500+ Walker humanoids delivered; targeting 10,000 units in 2026.
China vs. US Assessment
| China | US | |
|---|---|---|
| Unit volume | 4 of 5 top manufacturers | — |
| Price | BOM ~$46K | BOM ~$131K |
| Manufacturers | 140+, 330+ models | < 10 meaningful players |
| Government backing | Active procurement, 15th Five-Year Plan | Minimal direct support |
| AI model sophistication | Less documented | π0, Helix, GR00T more advanced |
| Commercial deployment | Scaling | More documented sustained deployments |
China has manufacturing-scale advantage. The US has AI-model advantage. Which one compounds faster over the next five years is the key strategic question.
Where It Goes from Here
The Disruption Sequence
Industry analysts broadly agree on the wave structure:
Wave 1 (2025–2027): Logistics and Warehousing — Already proven. Defined tasks, measurable ROI. 33% of projected installations.
Wave 2 (2026–2028): Automotive Manufacturing — Semi-structured. Figure/BMW and Atlas/Hyundai are bridgeheads.
Wave 3 (2027–2030): General Manufacturing — Broader task variety. Requires more general AI capability.
Wave 4 (2029–2032): Elder Care and Home — Requires safety certification, social acceptance, sub-$20K pricing, and sufficient dexterity for unstructured residential environments.
Wave 5 (2030+): Construction — Hardest environment. Most capable platforms required.
The “iPhone Moment” Timing
Consumer estimates cluster around 2027–2029:
- Tesla Optimus consumer availability targeted end-2027
- Foundation models reaching sufficient reliability for home environments
- Price crossing the $20K threshold
BofA projects humanoid ownership exceeding car ownership by 2060. Annual shipments hitting 1.2M by 2030 — 86% CAGR from 2026. Aggressive, but the direction is not contested.
NVIDIA’s Jensen Huang said a “ChatGPT moment” for robotics is imminent. He said that at CES in January 2025. The moment may be closer than even that framing suggested.
The 36-month transition from demo to deployment is behind us. The next 36 months will determine whether the hardware cost curve and the AI capability curve intersect fast enough to trigger the consumer wave. The pieces are moving.
Post 1 of the Autonomous Frontier series.