Intelligence
35%Autonomy in the wild: unseen-task success, zero-shot generalization, and safe operation around people — demonstrated, not demoed.
The Deployment Audit · Audit 001 — Q3 2026
The public scoreboard of the quarterly Deployment Audit. Robots are ranked only on verified work — logged hours, factory counts, trial results. Claims without evidence are shown as gaps, never as scores.
Board last updated Sep 30, 2026
Four vectors, one gate, one utility signal. The weights below were locked September 26, 2026. Every vector is scored 0–100 from published, attributed evidence — and a vector with no verified evidence is a gap, excluded from the grade rather than scored as zero.
Autonomy in the wild: unseen-task success, zero-shot generalization, and safe operation around people — demonstrated, not demoed.
Sustained operation: verified operating hours, battery and charging duty cycle, fleet uptime in real facilities.
Verifiable work capacity: payload the robot demonstrably handles, repeatedly, under load — not a brochure figure.
Movement capability: reach, dexterity, and locomotion across real terrain and real workstations.
Before any robot can be graded, it must clear the anti-vaporware gate: 100 hours of autonomous operation without an E-stop intervention.
One honest note: no manufacturer currently publishes E-stop intervention logs. Until they do, gate status is evaluated on fleet operating evidence — and marked as such on every card below.
The single signal the four vectors feed. A grade resists headline specifications that do not translate into dependable work. Composite grades require at least two vectors with verified inputs; grades resting on fewer than four vectors carry a partial-grade mark.
Only three builders cleared the bar for a graded entry this quarter: each has at least two vectors backed by verified, dated evidence. Everyone else is on the watchlist — real companies, ungraded until the receipts exist.
GATE FLEET EVIDENCE — 24/7 operation at CATL Ningde since Mar 2026 acceptance (company-reported). E-stop logs unpublished.
Gold in robot skills (pharma sorting), fully autonomous, no teleoperation — World Humanoid Robot Games, Beijing, Aug 2025.
Running 24 hours/day at CATL Ningde battery base since acceptance testing, Mar 2026. CNY 236M procurement contract.
Payload figures conflict between developer manual and launch PR. Shown as a gap — not scored, not zero.
Wheeled semi-humanoid mobile manipulator (not bipedal). 170+ autonomous retail units across 40+ cities (GalbotStore).
Evidence on file
What moves this grade: publish a single reconciled payload figure for Power, and the partial mark comes off. E-stop logs would clear the gate outright.
GATE FLEET EVIDENCE — 65,000+ commercial hours at Amazon, Toyota, GXO (company-reported). E-stop logs unpublished.
100,000 totes handled at ~98% accuracy at GXO Atlanta (company claim, May 2026). Fence-free cageless design on Digit 5 — certification pending third-party validation.
65,000+ commercial operating hours across Amazon, Toyota, GXO. Digit 5: 90-min battery, 9-min charge, ~20 hrs/day duty cycle (preliminary company specs, Sept 15, 2026).
50 lb (22.7 kg) payload on Digit 5, +40–43% vs Digit 4. Preliminary company specs based on pre-production hardware testing.
Bipedal locomotion, 2.2 m reach, interchangeable hands. 181 cm / 129 kg (preliminary company specs, Sept 15, 2026).
Evidence on file
What moves this grade: third-party certification of fence-free operation lifts Intelligence; published E-stop logs clear the gate outright.
GATE FLEET EVIDENCE — 11-month BMW Figure 02 deployment complete; Figure 03 on line since June 2026 (company-reported). E-stop logs unpublished.
Helix 2.5: 56% zero-shot completion across 420 long-horizon tasks in 30 unseen homes. Company-announced Sept 17, 2026 — no independent replication yet.
11-month BMW Figure 02 deployment complete; Figure 03 on the line since June 2026. Note: duration evidence is Figure 02, not 03.
No verified payload figure on file. Shown as a gap — not scored, not zero.
No verified motion-capability figure on file. Shown as a gap — not scored, not zero.
Evidence on file
What moves this grade: independent replication of the Helix results; verified Figure 03 payload and uptime figures. Company-announced is a receipt — independent verification is a better one.
No verified deployment yet — or not enough verified evidence to score two vectors. These are real companies with real programs. They enter the rankings the quarter their receipts arrive.
~15,000 units of parts ordered for 2026 (supply-chain reports); 1M units/yr Fremont run-rate target (Musk, official line); rare-earth magnet export-license exposure confirmed on Tesla’s April 2025 earnings call.
Supply-chain reporting · Sep 2026 / Musk official statementsNo verified commercial deployment. Early units route to Tesla’s internal “Optimus Academy” training operation — not sale, not logged hours, not a grade.
R1 official pricing $4,900 (Air) / $5,900 — the industry’s price-collapse benchmark. 20–26 DOF, ~1 hr battery. Real, purchasable, shipping; the research-lab default chassis.
Unitree official pricing · verified Sep 2026Ships units, not logged autonomous work. No verified deployment hours, no verified payload-under-load, no independent uptime record on file.
20,000 robots cumulative (company, Sept 2026); 10B yuan 2027 revenue target; ~100 hrs of factory livestream operations at a tablet plant.
AgiBot company reporting · Sep 2026Cumulative shipments are not logged operating hours. Vector-level evidence — per-task success rates, duty cycles, E-stop record — is insufficient for a grade.
Mercedes-Benz equity + factory pilots; Jabil manufacturing partnership; GXO partnership; ~$935M raised at $5B+ valuation; Apollo 2 unveiled June 30, 2026; <$50K BOM target.
CNBC · 2026 / Apptronik company announcementsPilots are not deployments. No verified deployment count, no logged hours, no task success figures published.
Robotics Metaplant Application Center opened inside Hyundai’s Georgia Metaplant (Sept 21, 2026) — Atlas training on factory logistics and parts sequencing. >25,000-unit deployment plan; 30,000/yr robot-factory capacity by 2028.
Hyundai / trade press · Sep 21, 2026 / JPMorgan investor session, May 2026The 25,000-unit figure is a company-stated plan — aspirational, not verified. No external customer named yet (expected early 2027). Reuters, Sept 14, 2026: no IPO in 2027.
100,000 units by late 2027 target (company plan); consumer home-humanoid program.
USA Today · May 1, 2026No verified customer home delivery as of July 2026. A target is not a deployment.
Guangzhou IRON production line commissioned (>80% automated); mass production targeted end-2026; China and overseas deliveries in 2027; $900M raise at $6.3B.
XPeng company announcements · Sep 8, 2026A commissioned line is not a delivery. No verified customer deliveries on file.
Bumi humanoid at ~$1,300 starting — price-collapse tier. Hobbs bionic-head platform with reported 32 DoF facial movement; heated-skin demo footage reported.
Noetix / demo reporting · 2026Heated-skin and full-facial-movement claims come from social/demo footage — reported, not lab-verified. No verified deployment of any kind.
Log 100 hours of autonomous operation without an E-stop intervention. The gate is a qualifier — it earns no points.
Operating hours, factory counts, trial success rates — with source and date. Company-reported counts, labeled as such.
Each vector is scored from verified inputs. At least two verified vectors are required for a composite grade.
Grades expire. Every quarter the Deployment Audit re-scores on fresh evidence — grades can fall as well as rise.
The honesty rule: a missing number is shown as a gap — never invented, never averaged in as zero, never smoothed over. If a grade is wrong because the evidence was wrong, the correction runs on camera and the record is updated in public.
If you can see the future, you’ve already won the future.
The NexaRaubotics mission