How rovers fail

The failure record holds 54 sourced entries from 15 teams plus the judges’ field-wide statements, set against 162 finals team-years. Forty-eight entries name a mechanism or a symptom inside a known subsystem and six are zeros whose cause was never published. The record does not say that comms is the problem in the way the judges describe it. It says that the largest and most persistent loss in the field is an autonomy capability gap, and that the most expensive single failures are the ones that take several missions down at once.

The database and its three biases

Every count on this page is a count of write-ups. Three biases shape the record and each is worth stating before a number is used.

Survivor bias in team reports. System acceptance videos are a qualification deliverable, so they show successes. Teams explain a failure mainly when it drove a redesign; a 2023 failure that ended a season and was never written up leaves no trace. The competition publishes no did-not-finish cause codes and no judging sheets.

Attribution bias in news releases. A release names the incident but rarely the mechanism. A “communication loss” is a symptom, not a cause, and where the team itself gave the mechanism, as WVU did for a damaged cable in 2026, that is the entry used S-WVU-NEWS2026.

A zero is not one thing. A zero can mean attempted and failed, did not attempt, or withdrew, and the score sheets do not distinguish them S-URC22-SCORES S-URC26-SCORES. Every causal reading on this page is therefore an inference rather than a measured cause.

The taxonomy

Counts are entries, not incidents: one bad mission can produce two, as Missouri S&T’s 2024 Delivery did with a failed signal mechanism and a failed arm axis. “Mechanism confirmed” means the source names the physical mechanism rather than only the score.

The count measures publication, not occurrence

The class with the most entries is not the class the judges name. The arm class and the comms class tie at seven. That is not evidence that the arm breaks as often as the link, because the judges state that comms is “far and away the most common problem that stops a team from being a top contender” S-URC-ADVICE. It is evidence of which failures get written up: a team explains a broken gearbox because it redesigned the gearbox, and explains a loose connector less often because the fix was a cable tie in a car park.

The same caution applies to the thin rows. Power has one sourced electrical failure in five years and dust has none at all, despite every subsystem study recommending sealed bearings, rated connectors and protected lenses. Two readings are possible and the data cannot separate them: those failures are rare, or they are fatal and under-published. Given that the judges describe rovers dying from a loose wire as routine, the second reading is at least as likely. The design consequence is to build to the mechanism rather than to the count.

Three further caveats belong with every percentage below. The database is a publication census rather than an incident census. Three teams account for 24 of the 54 entries because they are the best documented, and the zero rates are an upper bound on failure frequency and a lower bound on capability gaps.

The zero-score record

Multiplying each zero rate by the mean non-zero score in that mission, 27.9 for Delivery, 36.3 for servicing and 45.5 for autonomy, gives the field’s points forgone: about 945 mission points a year, of which about 610 are autonomy, 196 servicing and 140 Delivery S-URC22-SCORES S-URC23-SCORES S-URC24-SCORES S-URC25-SCORES S-URC26-SCORES. It is an upper bound, because a team that scored zero might have scored below the mean rather than at it.

Science has scored no zero since 2022, and the review element has never explained more than 5.7 % of the spread in nine years. The 300 points in Delivery, servicing and autonomy are where the field is decided, and all three are all-or-nothing for a substantial share of teams.

What the zeros are made of

If the three missions failed independently, servicing and Delivery zeros would coincide in about 2.6 % of team-years. The observed count is 12 of 162, or 7.4 %, 2.9 times the independent expectation.

Twenty-nine of 162 team-years had two or more zeros and eleven had three or more. Those are the rover-level failures, and they are where the one broken rover story lives. The isolated autonomy zero is the opposite case: the rover drove, serviced and sampled, then scored nothing when it had to drive itself.

That produces two failure economies with different investments. The blocks of 100 are drivetrain, power, connectors and the arm as a single point: they cost several missions at once and are bought off with redundancy, retention and test. The recurring 45 is autonomy, a capability 41 % of the field never demonstrates, bought with one simple planner that finishes. A first-season team that fixes the first and ignores the second lands near the median of about 230–260 points S-URC-RESULTS.

The decision behind each class, and the cheapest fix

Each row names the design decision that makes the class likely, not the component that broke. Cost is relative within this table only.

The root cause in one line: rovers fail where a single element is load-bearing and untested under field loads. The field under-tests exactly those elements, because testing them is boring, slow and happens in a car park rather than a laboratory.

The four load-bearing single points

The lift matrix says a servicing zero drags Delivery and autonomy with it, so reliability effort belongs on the elements that can zero three missions at once.

The FMEA tripwires that follow from these rows are FMEA-02, the connector, at 560; FMEA-06, pack sag, at 384; FMEA-05, spectrum congestion, at 336; FMEA-16, the arm zero, at 336; and FMEA-20, the one-metre arrival, at 294. FMEA-31, an untested feature shipped the week before the event, is also at 294 and is the row that creates the others.

Failure modes the 2027 rules add

No team has hit these yet, because the task did not exist in this form. They belong in the failure list before the field discovers them.

What a first build should do first

• The reliability budget belongs on the failure that zeroes three missions, not on the one that costs a task. The four single points above are the whole first season’s effort, and each is a small change compared with a redesign.
• The rover should be instrumented before the evidence is needed. No team publishes measured link, thermal or bus data. Signal strength, bus voltage at 100 Hz and enclosure temperature with an alarm turn every field day into data.
• Autonomy points are cheaper with simplicity than with sophistication. The cleanest natural experiment in the dataset is Monash’s autonomy score of 0 in 2024, 14 in 2025 and 86 in 2026, with the last step explicitly a simplification S-MONASH-SAR2025 S-MONASH-SAR2026. One planner, a progress checker, bounded recovery and ten cold-boot full runs beat a better algorithm.
• The transport leg is a mission of its own. The judges name it, Michigan found damaged parts on arrival and IIT Madras found a burned board S-UMICH-AERO-2026 S-IITM-URC2017-REPORT. A packaged shake test and a post-arrival check are the cheapest reliability measures available.
• The freeze date is written down before travel. Design-process entries appear at every level of the field, and three of the four largest single-year losses follow a whole-system change.
• Recovery is rehearsed as well as success. WVU’s requirement that each operator train at least 30 hours turned a damaged cable into 22.5 points instead of a zero S-WVU-NEWS2026.

Open problems

• The database is a census of write-ups. Fifty-four entries from 15 teams against 162 finals team-years in five years means absence of an entry is absence of documentation, not absence of failure.
• The zero-to-cause link is an association. The lift matrix cannot separate one broken rover from one team that had a bad weekend; a published intervention log would settle it in one table.
• Three teams supply 24 of the 54 entries, which is a documentation bias rather than a claim that those programmes fail more.
• No team publishes measured thermal, link or bus data, so the thermal and electrical classes are argued from mechanism rather than from field measurement.
• A new failure mode appears with each rule change, and the 2027 bioassay and astronaut tasks need wet rehearsal and a suited trial, neither of which can be arranged at short notice.

Further reading

• Mass and power carries the pack sag and the peak-current chain behind the power failures.
• Interfaces and couplings explains why a connector in the safety path is an interface rather than a component.
• Risk register turns these classes into a ranked register.
• Environment, dust and thermal covers the classes the field documents least.
• Missouri S&T, Monash and WVU hold the best-documented loss records.