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

Research library · updated 2026-06-15 · public

Figure / Unitree S4 Evidence Debt Matrix v1

Date: 2026-06-15 Owner: Hugo / Genius Team Agent: Finance / Charlie AGT-002 Status: public-safe research artifact Visibility: PUBLIC Primary use: morning review, /robotics/companies, slide 11 speaker notes, S4-to-S5 tracking

Related files:

  • knowledge/robotics-research-harness.md
  • knowledge/robotics-figure-unitree-s4-to-s5-conversion-dashboard-v1.md
  • knowledge/robotics-figure-history-product-deployment-timeline-v1.md
  • knowledge/robotics-unitree-deep-dive-v1.md
  • knowledge/robotics-figure-unitree-morning-review-evidence-audit-v1.md

Public-safety note:

  • Public-safe: yes.
  • No Hugo private portfolio data.
  • No trade recommendation.
  • No winner ranking.
  • This is an evidence-debt tracker: it identifies missing evidence, not investable conclusions.

0. One-line answer

截至 2026-06-15，Figure AI 和 Unitree / 宇树的公开证据已经足够支撑“机器人研究从 demo watch 进入 S3/S4 KPI tracking”，但还不足以支撑“S5 scaled-commercial economics”。Figure 的证据债务主要是 customer economics debt：robot count、contract value、ROI/payback、repeat order、uptime/intervention rate、service cost、gross margin。Unitree 的证据债务主要是 scale-and-economics debt：humanoid shipment volume、customer mix、industrial uptime、warranty/support burden、gross margin、software/service attach。🟢 official pages rechecked 2026-06-15; 🟠 Charlie evidence-debt classification.

1. What changed in this pass

This artifact does not add a new winner thesis. It adds a sharper morning-review tool:

  1. Turn Figure / Unitree comparison from “which company is better?” into “which evidence debt matters next?”
  2. Rank missing evidence by thesis impact.
  3. Separate S4 evidence already proven from S5 evidence still absent.
  4. Give Codex a compact public-safe module that avoids over-compressing deployment KPI and cost-access KPI into one score.

2. Source verification snapshot

As of 2026-06-15, direct web retrieval returned HTTP 200 for these official pages and matched the listed evidence anchors:

CompanyOfficial sourceAnchors matched on 2026-06-15Source gradeWhat this verifiesWhat it does not verify
Figure AIhttps://www.figure.ai/news/production-at-bmw90,000, 1,250, 30,000🟢 primary company sourceBMW deployment page remains reachable and contains key operating anchors.Contract value, robot count, ROI/payback, repeat order, intervention rate, gross margin.
Figure AIhttps://www.figure.ai/news/ramping-figure-03-production350, 80%, 99.3🟢 primary company sourceBotQ / Figure 03 manufacturing-KPI page remains reachable and contains production/yield anchors.Sell-through, customer utilization, revenue, margin, field failure rate.
Figure AIhttps://www.figure.ai/news/figure-signs-agreement-with-catalyst-brandsCatalyst🟢 primary announcement / 🟠 economics undisclosedCatalyst commercial-agreement page remains reachable.Robot count, contract value, deployment KPI, repeat order.
Unitreehttps://www.unitree.com/R14,900🟢 primary product pageR1 low-price anchor remains reachable.Shipment volume, reliability, gross margin, support cost.
Unitreehttps://www.unitree.com/g1/13.5K🟢 primary product pageG1 public price anchor remains reachable.Industrial deployment or customer economics.
Unitreehttps://www.unitree.com/H229,900🟢 primary product pageH2 public price anchor remains reachable.Product-line margin, availability, shipment scale.
Unitreehttps://www.unitree.com/H2plusJetson, GR00T🟢 primary product page / 🟠 platform implicationH2 Plus robotics-stack references remain reachable.Unitree-owned software revenue or generalized autonomy.
Unitreehttps://www.unitree.com/G1-Ddata acquisition🟢 primary product page / 🟠 platform implicationG1-D data/model workflow language remains reachable.Active developer count, paid software, dataset scale, customer ROI.

Method: simple HTTP retrieval with browser-like user agent on 2026-06-15; keyword presence check only. This verifies source reachability and anchor presence, not independent factual accuracy or commercial outcome. 🟠

3. Current evidence vs evidence debt

Company / pathCurrent strongest evidenceCurrent signalEvidence debtDebt severityWhy it matters
Figure customer-site deploymentBMW: 11-month deployment; 10-hour shifts Monday-Friday; 90,000+ parts loaded; 1,250+ runtime hours; contribution to 30,000+ BMW X3 vehicles. 🟢 Figure official post dated 2025-11-19; rechecked 2026-06-15.S4 deployment KPIRobot count by site; contract value; customer-confirmed ROI/payback; repeat order; uptime/intervention distribution.HighThis is the shortest path from “robot can work” to “customer repeatedly pays.”
Figure manufacturing rampBotQ / Figure 03: 350+ robots delivered; cadence from 1/day to 1/hour; >80% EOL first-pass yield; 99.3% battery-line first-pass yield; 9,000+ actuators in prior captured evidence. 🟢 Figure official post dated 2026-04-29; rechecked 2026-06-15.S4 manufacturing KPISell-through to customers; fleet utilization; production cost; field failure rate; gross margin; warranty/service burden.HighProduction-process KPI is valuable only if it converts into economically utilized fleet.
Figure customer expansionCatalyst commercial agreement starting at Reno, Nevada distribution logistics center. 🟢 Figure official post dated 2026-05-26; rechecked 2026-06-15.S3/S4 customer-surface signalRobot count; task KPI; deployment duration; contract value; rollout schedule; repeat expansion beyond first site.Medium-highTests whether BMW evidence generalizes across customer/task type.
Unitree cost-access curveR1 AIR from US$4,900 / R1 from US$5,900; G1 from US$13.5K; H2 US$29,900. 🟢 Unitree official pages accessed 2026-06-10; rechecked 2026-06-15.S3/S4 hardware accessHumanoid shipments by product line; customer mix; repeat orders; warranty/support burden; gross margin.HighLow price changes adoption surface but may either scale the market or compress hardware value.
Unitree developer/model workflowH2 Plus references Jetson / GR00T; G1-D references data acquisition and model workflow. 🟢 Unitree official pages rechecked 2026-06-15.S3 platform-enablementActive developer count; dataset scale; paid software/tooling revenue; model ecosystem adoption; deployments using workflow.MediumDetermines whether Unitree captures platform value or mostly sells commoditized bodies.

4. Evidence-debt priority order

P0 — evidence that would change the thesis fastest

  1. Figure customer ROI/payback or repeat order from BMW / Catalyst / third named customer. 🟢 needed.

    • Why P0: It directly connects deployment KPI to economic repeatability.
    • Current status: absent from reviewed public sources as of 2026-06-15. 🟢/🟠

  2. Figure robot count by customer site plus uptime/intervention distribution. 🟢 needed.

    • Why P0: Runtime hours without robot count and intervention profile cannot be converted into per-robot productivity.
    • Current status: absent from reviewed public sources as of 2026-06-15. 🟢/🟠

  3. Unitree humanoid shipments by model and customer type. 🟢 needed.

    • Why P0: Low price is only investable-relevant if it becomes real shipment scale and repeat demand.
    • Current status: absent from reviewed public product pages as of 2026-06-15. 🟢/🟠

  4. Unitree gross margin / warranty / support burden by humanoid product line. 🟢 needed.

    • Why P0: Low ASP can widen adoption while destroying hardware value capture; margin/support data separates these cases.
    • Current status: absent from reviewed public sources as of 2026-06-15. 🟢/🟠

P1 — evidence that strengthens or weakens S4 quality

  1. Catalyst deployment follow-through: robots deployed, task quantity, runtime, timeline, repeat expansion. 🟢/🟡 needed.
  2. BotQ output linked to contracted deployment, not only internal production/delivery count. 🟢/🟠 needed.
  3. Unitree named industrial customer KPI with runtime, task success, intervention, safety record and support burden. 🟢/🟡 needed.
  4. Unitree H2 Plus / G1-D developer adoption metrics: active users, paid tooling, dataset scale, model performance or deployment conversion. 🟢/🟠 needed.

P2 — useful, but not thesis-changing alone

  1. Additional demo videos without runtime/intervention/customer context. 🔴/🟠 noise unless upgraded.
  2. Additional product specs without shipment/economics. 🟢 for specs / 🟠 for commercial implication.
  3. More partnerships without deployment KPI or economics. 🟢 for announcement / 🟠 for economics.
  4. Third-party teardown or BOM estimates without customer/order validation. 🟡/🟠 support only.

5. How to use this in a public slide or site module

Suggested title

Figure / Unitree: stronger evidence, different debts

Public-safe message

Figure has stronger measured-work evidence: BMW runtime/task KPI and BotQ manufacturing KPI. Unitree has stronger cost-access evidence: sub-US$5k entry price, sub-US$30k higher-end humanoid anchor, and developer/data workflow packaging. Both are better than demo-only evidence. Neither is S5 until the debt is paid: customer ROI, repeat orders, low intervention, shipment scale, gross margin and support cost.

Three-card layout

  1. Figure: deployment debt

    • Proven: BMW 1,250+ runtime hours, 90,000+ parts, 30,000+ X3 contribution. 🟢
    • Missing: robot count, ROI/payback, repeat order, intervention rate.

  2. Figure: manufacturing debt

    • Proven: 350+ Figure 03, 1/day -> 1/hour cadence, >80% EOL FPY, 99.3% battery-line FPY. 🟢
    • Missing: sell-through, utilization, revenue, margin, service cost.

  3. Unitree: scale/economics debt

    • Proven: R1 from US$4.9k, G1 from US$13.5k, H2 US$29.9k; H2 Plus / G1-D workflow. 🟢
    • Missing: shipments, customer KPI, reliability, support cost, gross margin, paid software attach.

Required footer: Evidence map only. No winner ranking. No trade recommendation. Deployment KPI ≠ full economics; low price ≠ reliability or margin.

6. Signal vs noise

Signal

  • Named-customer operating KPI with duration, task quantity, runtime and production context. Figure BMW is S4. 🟢
  • Manufacturing KPI with cadence, yield and component output. Figure BotQ is S4. 🟢
  • Public humanoid price anchors that materially lower experiment cost. Unitree R1 / G1 / H2 are S3/S4 cost-access signals. 🟢/🟠
  • Data/model workflow packaging that could become a developer platform if adoption is later quantified. Unitree H2 Plus / G1-D are S3 platform signals. 🟢/🟠

Noise unless upgraded

  • Treating a customer logo or commercial agreement as economics without robot count, value, task KPI or repeat order. 🟢/🟠
  • Treating low price as reliability, gross-margin, shipment-scale or ROI proof. 🟠
  • Treating manufacturing output as economically utilized customer fleet. 🟢/🟠
  • Treating NVIDIA / Isaac / model workflow references as Unitree-owned software economics. 🟢/🟠
  • Treating viral videos as evidence without reset count, intervention rate, runtime, safety record and customer context. 🔴/🟠

7. What would change our mind

Figure upgrade toward S5 if:

  1. BMW, Catalyst or another customer confirms robot count, paid contract economics, repeat order, ROI/payback or multi-site rollout. 🟢
  2. Figure publishes long-duration uptime/intervention distribution tied to customer production use. 🟢/🟠
  3. BotQ output is connected to contracted deployments, backlog/revenue, field utilization, gross margin and service burden. 🟢/🟠
  4. Multiple customers show the same deployment playbook across manufacturing and logistics. 🟢/🟠

Figure downgrade if:

  1. BMW remains the only deeply quantified deployment after another 6-12 months. 🟠
  2. Catalyst remains announcement-only without robot count, deployment KPI or timeline. 🟠
  3. Production cadence rises while customer utilization, field reliability and revenue remain undisclosed. 🟠
  4. Service/support burden offsets production learning and customer productivity. 🟠

Unitree upgrade toward S5 if:

  1. Unitree discloses humanoid shipment volume by R1/G1/H2 line, customer type and repeat-order rate. 🟢
  2. Named industrial customers disclose uptime, task success, intervention rate, support burden and payback. 🟢/🟡
  3. Unitree discloses gross margin, warranty cost, maintenance cost or recurring service/software revenue for humanoid products. 🟢
  4. H2 Plus / G1-D becomes a measured developer standard with active developer count, dataset scale, paid tooling or deployment conversion. 🟢/🟠

Unitree downgrade if:

  1. Low-cost humanoids remain mostly demo / education / research hardware without reliable industrial workloads. 🟠
  2. Headline prices are offset by support, warranty, maintenance, safety or availability constraints. 🟠
  3. Value migrates to NVIDIA, open models, integrators or customers while Unitree captures mostly low-margin hardware. 🟠
  4. Public prices do not translate into shipment scale, repeat purchases or durable customer usage. 🟠

8. Common misconceptions

  1. Misconception: “Figure has BMW KPI, so humanoid commercialization is solved.”

    • Correction: BMW KPI is strong S4 measured-work evidence, but S5 needs customer economics: ROI/payback, repeat order, contract value, intervention rate, margin and service cost. 🟢/🟠

  2. Misconception: “Unitree is cheap, so Unitree wins.”

    • Correction: Low price lowers experimentation cost; it can also compress hardware margins. S5 depends on shipments, reliability, customer ROI, support burden and value capture. 🟢/🟠

  3. Misconception: “Manufacturing output equals commercial deployment.”

    • Correction: Produced/delivered robots, customer utilization, recognized revenue and gross margin are separate evidence units. Figure BotQ is S4 production-process evidence, not S5 economics. 🟢/🟠

  4. Misconception: “Developer workflow means software revenue.”

    • Correction: H2 Plus / G1-D workflow language shows platform surface, not paid software attach, active developer base or software gross margin. 🟢/🟠

9. Think Deeper questions

  1. If only one new Figure datapoint appears, which changes the thesis more: robot count, intervention rate, customer ROI/payback, repeat order, or gross margin?
  2. If Unitree shipment scale appears before gross margin, is that a value-capture signal or a commoditization warning?
  3. What is the right evidence discount when the KPI is company-disclosed rather than customer-confirmed?
  4. Which evidence would matter more for public-market proxies: OEM deployment economics, supplier revenue exposure, compute/software attach, or customer productivity capture?
  5. Could Figure and Unitree both be right but value migrate elsewhere: NVIDIA/model layer, integrators, safety/certification, or customer operations?

10. Source list

11. Public-safe flag

Public-safe: yes, if used as an evidence-debt matrix.

Do not include:

  • Hugo private portfolio data.
  • Private channel checks.
  • Paid-report excerpts.
  • Buy / sell / hold language.
  • Winner ranking.
  • Any claim that Figure has proven full scaled economics.
  • Any claim that Unitree low price proves industrial reliability, shipment scale, gross margin or customer ROI.