New evidence, corrections, reclassifications and methodology changes that affect how the record can be understood.
Public data
The public evidence is now available as eight joinable CSV tables
Records, claims, disclosures, sources, evidence history, deployment signals, coverage and metric definitions can now be downloaded as separate CSV tables from the Coverage page.
Stable record, claim, source and state IDs preserve joins across the tables instead of relying on display names.
Every table has a versioned file, a stable latest copy, documented columns, row counts and a SHA-256 fingerprint in the CSV manifest.
The CSV files use UTF-8, spreadsheet-safe cell handling and the same public/private boundary as the JSON export.
Public export 1.2 contains 340 rows across the eight tables; private candidate observations, workflow data and retained third-party source bodies remain excluded.
Source integrity
Eight source records gain durable verification metadata
All four sources behind Figure 02 × BMW and all four behind Unitree G1 × Fraunhofer IPA now have exact section or page locators, retrieval metadata and SHA-256 fingerprints.
The captured HTML and PDF response bodies are retained in a restricted internal archive outside the public web root and registered append-only in PostgreSQL.
Record pages and public export 1.1 expose stable artifact IDs and verification metadata, but do not redistribute third-party files, response headers or internal paths.
A changed source creates a new artifact instead of replacing the earlier capture; content changes still require editorial review before they alter a Reliability Record.
This is a two-record pilot, not a claim that every HumanoidUptime source has been archived or independently authenticated.
Coverage and public data
The initial research universe is now visible and reusable
A new Coverage page publishes the complete 30-program starting universe and distinguishes 13 full records, 16 deployment signals and one program with no qualifying public result at the cutoff.
The page states the evidence period, review period, languages, source types, search themes and selection bias instead of presenting the set as a market census.
A dated JSON export, stable latest copy, data dictionary and SHA-256 checksum make the public evidence reusable without exposing candidate observations, editorial priorities or raw submissions.
Source metadata now includes explicit fields for exact locators, archive references and content fingerprints; null values make clear where HumanoidUptime currently preserves only the live URL, access date and metadata fingerprint.
An external spot-check register is public and currently empty rather than implying review that has not happened.
Evidence classification
Deployment evidence is now separate from reliability evidence
Record cards no longer use one overall evidence badge for a mixed body of claims.
Deployment or test evidence now states who confirms the activity, while reliability evidence separately states the strongest support for operating results.
Digit × GXO therefore shows an operator-confirmed deployment alongside manufacturer-reported output, rather than the ambiguous label ‘Mixed public record’.
An operator-confirmed pilot no longer implies that the operator also published reliability performance.
Methodology
Every record now has a defined system and mission boundary
HMND 01 × Siemens identifies one described robot as the unit of analysis and separates the known tote workflow and factory environment from incomplete schedule, success, energy and human-support boundaries.
Unitree G1 × Fraunhofer identifies the exact EDU-4 test asset, hands and firmware and keeps its laboratory mission separate from a production deployment.
Mission fields distinguish known, partially defined, not disclosed and narrowly reasoned not-applicable context without changing evidence provenance.
All 13 public records now have a stable analysis scope and complete mission profile; every claim is linked to that scope or marked as reasoned adjacent context.
Eight claims about another product version, site, stage or program remain visible as context but cannot supply evidence for the bounded record.
The two detailed profiles remain presentation pilots while record-page UI is reviewed as a separate step.
Methodology
Time claims are now tested before they become availability
All 13 records now use explicit Measurement Model v1 disclosure states; the review retains one narrowly reasoned not-applicable field and does not use N/A as a substitute for missing data.
Eight claims with potentially confusing runtime, calendar-duration, endurance, continuous-operation or uptime wording now preserve the source metric while separately testing whether availability can be calculated.
HMND 01 × Siemens keeps the published ‘uptime exceeding 8 hours’ wording but shows that scheduled time, downtime, scope and exclusions are missing, so no availability percentage is calculated.
Unitree G1 × Fraunhofer keeps derived per-charge endurance as an energy result rather than recasting it as availability.
‘Not applicable’ is now separate from evidence strength and requires a record-specific reason; field availability for the Unitree remains unknown outside the bounded laboratory benchmark.
Editorial governance
Accountability and review boundaries made public
A new About page identifies who is accountable for HumanoidUptime and discloses the project's current maturity, funding and conflicts without turning every record into a personal byline.
The Editorial Policy documents source discovery, corpus boundaries, language coverage, publication gates, corrections and the distinction between AI assistance and evidence.
Every Reliability Record now shows that it received human editorial review and links to the shared policy; HumanoidUptime does not claim universal two-person or external review.
Legal and Privacy pages now make operator contact and the site's actual access-log, rate-limit and evidence-submission processing visible.
New evidence
Walker S2 at ROSSMANN becomes Record 013
ROSSMANN confirms receiving one Walker S2 in Burgwedel on 17 March for a multi-phase logistics pilot spanning 2026.
UBTECH later reports material handling and sorting and makes a qualitative stability claim without definitions or measurements.
The planned calendar window is not treated as runtime, and Terra Robotics' wider purchase is not attributed to the ROSSMANN fleet.
Walker S1 × Foxconn, Persona × HD Hyundai and Apollo 2 × Robot Park remain source-audited signals rather than full records because their deployment evidence is not sufficiently bounded.
New evidence
Humanoid pre-alpha at Schaeffler becomes Record 012
Schaeffler corroborates a successfully completed proof of concept, and Humanoid identifies the Erlangen site and bearing-ring bin-picking workflow.
Humanoid documents leader-arm teleoperation for training and describes autonomous regrasp correction without publishing test-level autonomy or recovery results.
Later Alpha and Beta deployment plans remain separate from the evaluated pre-alpha configuration.
No test duration, task count, cycle time, throughput, intervention, failure, recovery or availability result was identified.
Coverage expansion
Five additional deployment candidates receive full evidence reviews
Figure × Catalyst Brands and Walker S2 × Hitachi remain manufacturer-reported integrations without a bounded operating result.
Walker S1 × BYD preserves claims of doubled efficiency and 30% higher stability, but neither measure has a published definition, baseline, sample or period.
IRON's September 2025 factory-training statement is kept separate from XPENG's later Next-Gen IRON product and commercialization roadmap.
Digit × Mercado Libre adds a named San Antonio fulfillment site, while Digit's earlier 100,000-tote figure is not attributed to that program.
No new full Reliability Record was opened because none of the five sources meets the publication gate.
New evidence
Atlas prototype at HMGMA becomes Record 011
Hyundai confirms that the Atlas prototype completed pilot testing at Hyundai Motor Group Metaplant America in October 2025.
The operator identifies the parts-sequencing workflow: picking automotive parts and aligning them in another rack.
HumanoidUptime keeps the tested prototype separate from the product model, planned 2026 RMAC fleets and Hyundai's 2028 production roadmap.
No pilot duration, fleet size, task count, autonomy share, intervention, failure, recovery or availability result was identified.
Deployment evidence
Apollo at GXO added with its lab-to-field boundary intact
GXO's June 2024 announcement establishes an early-stage laboratory proof-of-concept, not a warehouse deployment result.
Apptronik's June 2026 update says Apollo 2 data-collection workflows are present at GXO customer sites and use teleoperation plus autonomous execution.
The public deployment directory now shows the later phase as manufacturer-reported because no dated GXO source confirms the specific site-level workflow.
No site, fleet size, operating hours, task results, autonomy share, intervention rate, failure or recovery result was identified.
Evidence history
Historical baseline audit completed
All ten existing records were checked once for source-bounded historical reconstruction; five qualified and five were deliberately left without history cards.
Unitree G1 × Fraunhofer now shows the path from independent safety measurements to the broader benchmark and detailed operating-boundary results.
HMND 01 × Siemens shows quantified manufacturer reporting later corroborated by Siemens, while Apollo × Mercedes-Benz shows the move from agreement to operator-confirmed factory training.
The baseline now contains twelve effective historical states across five records; future history will be created forward through normal evidence reviews.
Evidence history
Public knowledge can now be followed over time
Historical evidence states now separate the date evidence became public from the later date HumanoidUptime reconstructed it.
Figure 02 × BMW shows the progression from a 2024 operator-confirmed trial to 2025 manufacturer metrics and 2026 operator corroboration.
Digit × GXO shows the transition from proof-of-concept to a commercial agreement without treating commercial maturity as reliability disclosure.
Reconstructed states are explicitly labelled, source-linked and append-only; they describe public knowledge rather than inferred changes in robot performance.
The history audit corrected Figure 02 autonomy from operator-corroborated to manufacturer-reported because the autonomy-specific wording in BMW's release is attributable to Figure; runtime and output remain operator-corroborated.
Explanation layer
Human help and autonomy explained
A new guide separates teleoperation, supervised autonomy and unattended task execution without treating autonomy as a universal robot label.
Supervision, planned input, intervention and recovery are distinguished so human presence is not automatically counted as assistance.
A worked example shows why completed output can differ from autonomous completion and intervention rates.
Autonomy and intervention fields plus relevant glossary terms now link directly to the deeper guide; task attempt now has its own stable glossary definition.
New evidence
Apollo production-environment training
Mercedes-Benz confirms that Apollo robots collected data at its Berlin-Marienfelde Digital Factory Campus for production use cases.
Employees transferred production knowledge through teleoperation and augmented reality before further autonomy training.
A later Apptronik update places Mercedes-Benz within broader mixed teleoperation and autonomous data-collection workflows without publishing a site-specific result.
No public operating hours, task output, fleet count, autonomy share, intervention frequency, downtime, failure or recovery result was identified.
New evidence
AEON factory-pilot milestones
BMW's operator record establishes the path from laboratory evaluation to a December 2025 test at Plant Leipzig and staged pilot integration in 2026.
Hexagon reports that AEON began production tasks on 12 June while continuing training for future manufacturing applications.
Battery assembly and component production remain bounded by the stages described in each source rather than being treated as completed output.
No public runtime, scheduled time, task count, autonomy share, intervention, failure, recovery or fleet-size result was identified.
Explanation layer
Reliability concepts now have a plain-language learning path
A new Explained hub introduces the measurement chain behind humanoid reliability evidence.
Two first guides show how to read a record and why operating hours cannot be converted into uptime without a defined denominator.
A 32-term glossary covers reliability, autonomy, evidence and common worksite vocabulary.
Reliability fields and selected record terms now open contextual definitions with links to deeper guidance.
Explanation layer
Every record now includes a plain-language conclusion
Each record states what is known, what remains unknown and why the distinction matters without assuming reliability-engineering knowledge.
The concise explanation is shown before the claim-level evidence, technical fields and sources.
Plain language does not introduce estimates or conclusions beyond the underlying evidence assessment.
New evidence
Digit factory-use evidence
Digit's Cheraw factory task is traced across operator, manufacturer and independent-reporting sources.
The sources establish a public activity span of more than one year, but not cumulative runtime or measured availability.
The named Cheraw workflow is confirmed without treating a global purchase agreement as evidence of a delivered fleet.
The reviewed sources still disclose no cumulative runtime, downtime, intervention or recovery record.
Methodology change
Evidence states made easier to read
Independent measurement, operator corroboration, manufacturer reporting, capability claims and undisclosed fields now use distinct evidence marks.
Missing reliability data is shown as an explicit finding instead of an ambiguous blank state.
Disclosure matrices and record details include accessible evidence-level text.
Every public record has a durable HU-R identifier.
Record review
Evidence boundaries clarified across five records
The Figure 02 × BMW record now shows the unresolved difference between BMW's ten-month scope and Figure's eleven-month description.
The Digit × GXO record distinguishes operator evidence and the joint commercial announcement from Agility-only output, duration and autonomy claims.
Walker S2 battery-swap autonomy is kept separate from task autonomy, and Figure 03 project language from measured field operation.
The Unitree G1 1:49-hour result is classified as scenario-derived endurance and includes Fraunhofer's collision-test repeat context.
New evidence
Historical retail pilot
A one-week Sanctuary AI × Canadian Tire/Mark's retail pilot is now part of the public record.
The manufacturer reports 110 correctly completed tasks but provides no denominator for a success rate.
The robot generation and autonomy mode remain unknown; the later Phoenix name is not applied retroactively.
New evidence
Factory POC evidence
Three first-party sources establish the HMND 01 Alpha × Siemens factory proof of concept.
The reported eight-plus uptime hours remain separate from availability, MTBF and continuous autonomous runtime.
Siemens corroborates throughput and autonomous pick-and-place results, while key denominators remain unavailable.
Coverage expansion
Deployment signals separated by maturity
Field programs, pilots, plans, benchmarks and product claims are shown as distinct stages.
Every full record distinguishes what is established, claimed and still unknown.
The comparison view explains why superficially similar public metrics may not be comparable.
New evidence
First independently measured record
Fraunhofer IPA's Unitree G1 EDU-4 benchmark becomes the first independently measured record.
The record covers measured energy, endurance, movement, safety, thermal and security findings with test context.
Per-charge endurance remains separate from cumulative field runtime.
Initial publication
Initial public evidence board
The public record begins with three source-linked cases and a reliability disclosure matrix.
Each case separates reported results, claim provenance and missing reliability fields.
The governing rule is: We don't score what nobody has measured.