Investment ResearchTopic research, evidence curves, and cycle signals
HomeSign inEN

Research library · updated 2026-06-13 · public

Figure AI / Unitree public compression guardrail v1

Date: 2026-06-13 Owner: Hugo / Genius Team Agent: Finance / Charlie AGT-002 Status: public-safe Codex packaging guardrail Visibility: PUBLIC Primary use: slide 11 / /robotics/companies / morning-review speaker notes

Related files:

  • knowledge/robotics-figure-unitree-evidence-v1.md
  • knowledge/robotics-oem-three-evidence-curves-public-brief-v1.md
  • knowledge/robotics-public-claim-bank-v1.md
  • knowledge/robotics-research-harness.md

0. One-line answer

截至 2026-06-13，Figure AI 和 Unitree / 宇树不应该被包装成“谁领先”的 winner ranking；它们是两条不同证据曲线：Figure 的强项是客户现场 deployment KPI 与制造 KPI，Unitree 的强项是低成本硬件 access 与 developer / model workflow 表面。两者都还没有公开披露完整 S5 scaled-commercial-economics proof。🟢 primary company sources captured in robotics-figure-unitree-evidence-v1.md, accessed 2026-06-10; 🟠 Charlie synthesis as of 2026-06-13.

1. Core packaging problem

Codex 压缩 slide/site 时最容易犯的错误是把三类不同问题合并成一个分数：

  1. Figure 的 BMW runtime / task KPI answer: “humanoid 能不能在真实工业场景做可计量工作？”🟢
  2. Figure 的 BotQ manufacturing KPI answer: “robot production process 是否从 hand-built prototype 走向可重复制造？”🟢
  3. Unitree 的 price / platform answer: “humanoid hardware access cost 是否低到足以扩大实验基数？”🟢

这些都不是同一个问题，也都不等于：customer ROI、repeat order、gross margin、low intervention、service cost、recognized revenue。🟠

2. Compression-ready evidence table

Evidence curveCompanyPublic evidence anchorSource grade / as-ofSignal gradeWhat it provesWhat it does not prove
Customer-site deployment KPIFigure AIBMW Spartanburg Figure 02 deployment: 11 months; 10-hour shifts Monday-Friday; 90,000+ parts loaded; 1,250+ runtime hours; contributed to 30,000+ BMW X3 vehicles.🟢 Figure official post dated 2025-11-19; captured 2026-06-10S4A humanoid system can be measured in a named industrial production context.Contract value, robot count, customer-confirmed ROI/payback, intervention rate, repeat-order economics, gross margin.
Manufacturing KPIFigure AIBotQ: 350+ Figure 03 delivered; production cadence improved from 1/day to 1/hour; EOL first-pass yield >80%; battery-line FPY 99.3%; 500+ battery packs; 9,000+ actuators.🟢 Figure official post dated 2026-04-29; captured 2026-06-10S4Production-process evidence is becoming measurable, not only capacity intention.Sell-through economics, utilization, field reliability, revenue, margin, service burden.
Commercial expansion signalFigure AICommercial agreement with Catalyst Brands to deploy humanoids in logistics network, starting at Reno, Nevada Distribution Logistics Center.🟢 Figure official post dated 2026-05-26; captured 2026-06-10; 🟠 economics undisclosedS3/S4A second named customer/application path beyond BMW-style manufacturing.Robot count, contract value, deployment schedule, ROI/payback, repeat order.
Hardware cost / access curveUnitree / 宇树R1 from US$4,900; G1 from US$13.5K; H2 US$29,900, tax/shipping excluded.🟢 Unitree official product pages captured 2026-06-10S3/S4Humanoid hardware experimentation cost is falling and public price anchors exist.Industrial reliability, shipment scale, gross margin, customer ROI, support cost.
Developer / model workflow surfaceUnitree / 宇树H2 Plus references NVIDIA Jetson T5000, FP4 2,070 TFLOPS, Isaac GR00T / TeleOp / Sim; G1-D describes data acquisition, labeling, model training, inference workflow.🟢 Unitree official product pages captured 2026-06-10S3Unitree is packaging more than robot bodies: compute, teleoperation, simulation, and data/model tools.Software revenue, developer adoption scale, model moat, deployment economics, Unitree capture of AI-layer economics.

3. Slide-safe headline options

Use one of these; avoid winner language.

  1. “Figure vs Unitree: deployment proof vs cost-curve proof.”
  2. “Two evidence curves, one missing proof: customer economics.”
  3. “Figure measures work; Unitree lowers access cost; neither yet proves S5.”
  4. “Do not score them the same way: runtime KPI is not price access, and price access is not ROI.”

Recommended title for slide 11:

“Figure / Unitree：两条证据曲线，不是一场简单 horse race”

Recommended subtitle:

“Figure = deployment + manufacturing KPI；Unitree = hardware cost + developer-platform access；S5 还需要 repeat paid deployment、ROI/payback、low intervention、margin。”

4. Three-card slide layout

Card 1 — Figure: measured work, economics missing

Approved copy:

Figure 的公开强信号不是 demo，而是 BMW deployment KPI 与 BotQ manufacturing KPI：BMW 披露 11 months、10-hour shifts Monday-Friday、90,000+ parts loaded、1,250+ runtime hours、30,000+ BMW X3 vehicles contribution；BotQ 披露 350+ Figure 03、1/day to 1/hour、EOL FPY >80%、battery-line FPY 99.3%。🟢 Figure official posts, 2025-11-19 and 2026-04-29.

Footer caveat:

Still missing: contract value, robot count by customer, customer-confirmed ROI/payback, intervention rate, repeat orders, gross margin. 🟠 as of 2026-06-13.

Card 2 — Unitree: lower access cost, reliability/economics missing

Approved copy:

Unitree 的公开强信号是 cost/access curve：R1 from US$4,900、G1 from US$13.5K、H2 US$29,900；H2 Plus / G1-D 把 robot body 接到 Jetson / Isaac / TeleOp / Sim / data-training workflow。🟢 Unitree official product pages captured 2026-06-10.

Footer caveat:

Still missing: humanoid shipments by product line, named industrial deployment KPI, uptime/intervention, warranty/support cost, gross margin, customer ROI/payback. 🟠 as of 2026-06-13.

Card 3 — What upgrades either to S5

Approved copy:

S5 不是更多视频，而是一条经济性证据链：paid deployment -> repeat order -> runtime / intervention -> maintenance cost -> customer payback -> production economics -> financial materiality。🟠 Charlie framework based on research harness, as of 2026-06-13.

Footer caveat:

Evidence map only. No winner ranking. No trade recommendation.

5. Do-not-say list and replacements

Do not sayWhy notSafer replacement
“Figure 已经证明 humanoid 商业化。”BMW / BotQ data are strong but economics are not disclosed.“Figure has among the strongest public deployment/manufacturing KPI evidence, but not yet full S5 economics.”
“Unitree 低价说明它会赢。”Low price may expand access or compress hardware margins; it does not prove reliability or margin.“Unitree creates a cost/access curve that may widen experimentation; value capture still depends on reliability, software/data, service, and economics.”
“Figure 和 Unitree 可以直接打分排名。”They answer different questions: measured work vs cost access.“Compare by evidence curve, not by one aggregate winner score.”
“350+ Figure robots means commercial sell-through.”The post says delivered robots / production metrics, but not customer sell-through economics.“350+ robots is manufacturing-process evidence; deployment utilization and revenue remain separate checks.”
“H2 Plus / G1-D means Unitree owns AI/software value.”Toolchain alignment does not prove paid software revenue or model moat.“H2 Plus / G1-D show platform-enablement surface; software/data value capture remains open.”
“Catalyst agreement is equivalent to scaled deployment.”Starting location is disclosed, but count/value/ROI/repeat terms are not.“Catalyst is S3/S4 commercial-expansion evidence until deployment economics are disclosed.”

6. Signal / noise rules for this comparison

Signal

  • Named customer plus quantified runtime/task output: Figure BMW = S4. 🟢
  • Manufacturing-process metrics with yield/cadence/output: Figure BotQ = S4. 🟢
  • Published humanoid price anchors below US$30K: Unitree R1/G1/H2 = S3/S4 cost-access signal. 🟢
  • Toolchain integration around compute, teleoperation, simulation, data/model workflow: Unitree H2 Plus / G1-D = S3 platform signal. 🟢
  • Commercial agreement with named initial site but undisclosed economics: Figure Catalyst = S3/S4. 🟢/🟠

Noise unless upgraded

  • Viral robot videos without runtime, task repetition, or intervention data. 🔴/🟠
  • Specs-only comparisons without customer deployment context. 🟢/🟠
  • Low price treated as proof of reliability or sustainable margin. 🟠
  • Customer logo treated as recognized revenue or repeat-order proof. 🟢/🟠
  • Production/cadence metric treated as field utilization. 🟢/🟠

7. What would change our mind

Figure upgrade path

Figure moves closer to S5 if BMW / Catalyst / another named customer confirms:

  1. robot count by site;
  2. paid contract value or recurring economics;
  3. uptime and intervention rate over long-duration production use;
  4. customer-confirmed ROI/payback;
  5. repeat orders or multi-site rollout;
  6. Figure revenue, gross margin, service/support burden tied to deployments.

Source requirement: 🟢 customer/company official disclosure or filing; 🟡 credible secondary only as triangulation; 🟠 estimate must show method.

Unitree upgrade path

Unitree moves closer to S5 if it discloses or customers validate:

  1. humanoid shipments by product line and customer type;
  2. industrial deployment KPI, runtime, and intervention rate;
  3. repeat orders / fleet expansion;
  4. warranty, maintenance, and support burden;
  5. gross margin / software-service attachment;
  6. developer ecosystem adoption with paying software/model/tooling evidence.

Source requirement: 🟢 official company/customer disclosure; 🟡 credible secondary only as triangulation; 🟠 estimates must state method.

8. Common misconceptions

  1. Misconception: “Deployment KPI automatically equals unit economics.” Correction: KPI can prove task performance, but economics need contract value, cost, intervention, maintenance, payback, repeat order, and margin. 🟠

  2. Misconception: “Low price automatically expands profit pool.” Correction: Low price can expand experimentation while compressing hardware value capture; profit may migrate to data, software, integration, service, or customer workflow ownership. 🟠

  3. Misconception: “Manufacturing cadence equals utilized fleet.” Correction: Factory output is a supply-side metric; field utilization is a demand/economics metric. Both are needed before S5. 🟠

  4. Misconception: “AI stack references imply software moat.” Correction: Jetson / Isaac / TeleOp / Sim references show ecosystem alignment; they do not prove proprietary model advantage or paid software revenue. 🟢/🟠

9. Think Deeper questions

  1. If Figure proves task KPI before full economics, what minimum extra disclosure would convert S4 into investable S5 evidence: ROI, repeat order, intervention rate, or gross margin?
  2. If Unitree pushes humanoid hardware toward sub-US$5K to sub-US$30K price bands, does the value pool migrate away from robot bodies and toward software/data/deployment?
  3. Which evidence should lead the sector’s next re-rating: repeat customer deployments, production yield, shipment volume, customer payback, or humanoid-specific revenue?
  4. Can Figure’s vertically controlled deployment route and Unitree’s accessible platform route converge, or are they building different markets?

10. Public site draft section

Figure / Unitree: compare evidence type, not hype level

Figure and Unitree are useful because they make two different parts of the humanoid robotics thesis measurable.

Figure is closer to the deployment question. Its BMW Spartanburg post disclosed 11 months of deployment, 10-hour shifts Monday-Friday, 90,000+ parts loaded, 1,250+ runtime hours, and contribution to 30,000+ BMW X3 vehicles. Its BotQ production post disclosed 350+ Figure 03 robots, cadence moving from 1/day to 1/hour, EOL first-pass yield above 80%, and battery-line first-pass yield of 99.3%. These are stronger than demo evidence, but still leave the economics question open.

Unitree is closer to the cost/access question. Its public product pages list R1 from US$4,900, G1 from US$13.5K, and H2 at US$29,900. H2 Plus and G1-D also attach the robot body to compute, teleoperation, simulation, data acquisition, and model-training workflows. This can widen the experiment base, but it does not prove industrial reliability, margin, support cost, or customer ROI.

The safe conclusion is simple: Figure measures more real work; Unitree lowers access cost. The next proof is not another impressive video. It is repeat paid deployment, low intervention, customer payback, production economics, and financial materiality.

11. Source list

12. Public-safety flag

Public-safe: yes.

No Hugo private portfolio data. No private-channel information. No trade recommendation. No winner ranking. Treat official company claims as primary for what the company says, but require customer-confirmed economics before upgrading to S5.