How robots fail and get fixed

A well-run industrial arm goes tens of thousands of hours between failures; an illustrative humanoid goes a few thousand, and a service robot's figure is not public at all. Warranties cover manufacturing defects and exclude the damage robots actually suffer, such as falls, collisions and spilled liquid. The maker holds the parts, the diagnostic software and the logs, so whoever fixes the robot works on the maker's terms. Those three facts decide whether a repair business in Southeast Asia can exist.

Three numbers: MTBF, MTTR and availability

Every service contract, every uptime guarantee and every insurance quote reduces to three numbers. A failure is any event that stops the robot doing its job until someone intervenes. The failure rate is how often that happens.

There is one trap in that formula. MTBF counts hours the robot runs; MTTR counts hours on the calendar. A robot that works one 10-hour shift five days a week and waits 20 days for a part loses 20 days of shifts, which is about 143 scheduled hours, rather than 480. The calculator below converts calendar repair time into scheduled hours lost, which is how a venue feels it.

What 99% uptime means in hours

Real claims to hold against the table: Formic, which rents factory robots by the hour, reported 100,000 production hours at 99%+ uptime in Jun 2024 verified. Weston Robot in Singapore claims 99.4% fleet uptime vendor claim. Industrial buyers expect about 99.9%, while humanoid task success drops from about 95% in the lab to about 60% in the field estimate. That gap is a reliability problem as much as an AI problem.

The bathtub curve

Failures do not arrive evenly over a machine's life. Plot failure rate against time in service and you get the bathtub curve: a high rate at the start while design flaws and installation mistakes shake out, a low flat floor of random failures, then a rising rate as parts wear out.

Where wear-out does arrive, makers already watch for it. Fanuc's ZDT (Zero Down Time) service connects 40,000+ arms, has prevented 4,000+ downtime cases, and detects reducer (gearbox) wear from motor torque "several weeks before actual failure" verified Jul 2025. That is predictive maintenance, run by the maker through its own cloud, on a data stream a third party with a laptop never sees.

How robots actually fail

An FMEA (failure mode and effects analysis) lists, for each part of a machine, how it can fail, what that does, how you would notice and how you fix it. The table below is FMEA-style: it uses field evidence from maker manuals and owner reports instead of scored ratings, because no maker publishes failure frequencies for service robots or humanoids.

Read the right-hand column and a pattern appears. For wheeled service robots, most "failures" are hygiene and consumables that venue staff can handle: fouled sensors, tangled wheels, worn squeegees, tired batteries, lost localisation. Real hardware repair is a module swap, and modules come from the maker. For humanoids, the most common failures in public reports are induced by control software, and they show up as heat and broken small joints.

Humanoids in the field

• Figure 02 at BMW Spartanburg: 10-hour shifts Monday to Friday, 1,250+ hours of runtime, 90,000+ parts loaded, 30,000+ X3 cars, about 1.2M steps, and more than 99% placement success per shift verified Nov 2025. Figure reports "minimal hardware failures" and names the forearm as the top failure point. The fleet retired after about 11 months when Figure 03 launched. Whether the 1,250 hours is per robot or for the fleet is not stated: if per robot, it ran about half of roughly 2,350 scheduled shift-hours contested.
• Agility Digit: 65,000+ operating hours at customers including GXO, Schaeffler, Amazon and Toyota Canada, and 100,000+ totes moved at one GXO site since Jun 2024 verified Sep 2026. Digit 5 claims more than 20 hours a day from a 90-minute battery and a 9-minute recharge claim.
• Apptronik: about 22 hours a day of uptime under RaaS, "not yet validated at fleet scale" unverified.
• UBTech Walker S2: autonomous battery swap in about 3 minutes, trials at BYD and Zeekr; no public failure or uptime data verified.
• Unitree G1 owners: overheating shoulder motors with a heavier dexterous hand, broken wrist gears after violent learned policies, sudden shutdowns when a policy switch trips battery over-current protection verified. One lab could not send its robot back because of a research deadline and asked whether it could open the actuator itself.

Warranties cover what rarely breaks

A warranty is the maker's promise to fix manufacturing defects free for a warranty period. An extended warranty buys a longer period. Read the exclusions and you find the business.

Three consequences. First, every excluded failure is paid repair, whoever does it, so demand for paid repair is structural. Second, the "shell opening" clause means an independent repairer cannot touch an in-warranty Pudu without voiding it: the in-warranty market belongs to whoever holds the distributor or authorised service contract. Third, Pudu sells extended warranties only while the original warranty runs verified, so the out-of-warranty base grows every month.

What warranties earn: US electronics sellers paid claims of 21% of recognised extended-warranty revenue by 2005 verified, old, and providers quote gross margins of 40 to 70% estimate. Those are consumer electronics that rarely fail. Restaurant and mall robots fail far more often, so treat those margins as a ceiling until someone measures robot failure rates.

How service is sold

After-sales is everything after the sale: repair, servicing, parts and updates. Chinese contracts call the maintenance-and-repair bundle maintenance and warranty service. The models below differ in who carries the risk of downtime.

Who delivers it matters more than the model. In Southeast Asia, distributors do the on-site work. Pudu's own page says "Distributors provide maintenance services to end-users, and Pudu Robotics provides free maintenance spare parts and technical support"; after warranty, the distributor pays for parts verified. Those distributors are small: Singapore's authorised Pudu distributor answers email at a coffee company's domain, Malaysia's was founded in 2022, and none publishes an SLA verified. The service level agreement a venue can actually get depends on that one company.

Parts logistics

Spare parts are the chokepoint. No maker or Southeast Asian distributor publishes parts lead times verified absence. The published anchors are Unitree's 20 days after it receives the broken part, plus shipping both ways at the customer's cost, and Pudu's and Gausium's rules that replaced warranty parts belong to the maker and must go back. Unitree may scrap damaged battery packs. Pudu tells owners to store robots above 60% charge and top them up every 3 months verified.

Two kinds of stock matter. Wear parts (squeegees, brushes, casters, wheels) are cheap, predictable and fit a van. Modules (lidar, drive wheel assembly, battery pack, mainboard, joint actuator) are expensive, brand-specific and only come from the maker. Repair prices for these modules in China and Southeast Asia are not public. Batteries also ship under dangerous goods rules (see compute, wiring and power), and customs times for parts are not in the research. A local spares kit is the obvious fix, and its cost is capital tied up per brand per country.

Who holds the keys

Diagnostic software, firmware, fault logs and parts sit with the maker and its authorised service providers. Pudu runs a formal partner channel with its own Service Partner Platform, next to separate platforms for distributors and venues, and describes its service tools as "original factory diagnostic systems" verified. Weston Robot is Unitree's authorised Global Service Center for Singapore and Malaysia verified. For an independent repairer, the realistic way in is to become such a partner, which makes it a subcontractor on the maker's terms.

Right to repair, status Sep 2026

In Southeast Asia, your access to parts, diagnostics and logs depends entirely on the contract you sign with the maker or its distributor.

A failure log record, field by field

Here is what one record from a repair visit could look like. It is illustrative: the fields are the point.

{
  "robot":    {"type": "tray delivery robot", "serial": "SN-0412", "firmware": "v3.2.1", "hours": 6120},
  "site":     {"type": "restaurant", "city": "Kuala Lumpur", "floor": "tile, wet near kitchen"},
  "fault":    {"time": "2026-09-19T21:40+08:00", "code": "LOC-02", "symptom": "lost, stopped at table 12"},
  "response": {"reported": "21:42", "remote_triage": "21:55", "on_site": "next day 10:30"},
  "fix":      {"cause": "lidar window fouled with grease", "action": "cleaned, recalibrated", "part": null, "minutes": 25},
  "downtime_min": 770,
  "sensor_snapshot": "held in the maker's cloud, absent from this record"
}

The last row is the problem with selling failure data to AI labs. Physical Intelligence's π*0.6 learns from expert corrections during its own robots' autonomous runs, and Dyna's DYNA-1 uses a reward model that watches robot video and targets error-recovery data verified. Every one of those loops runs on the lab's own robots, with synchronised camera frames, joint states and actions captured at the moment of failure. A repair ticket that says "left wheel tangled, 21:40, table 12" carries none of that, and no public price for third-party failure logs exists verified absence. The buyers who do need the other rows are the people who price risk. China's insurers have insured 1,000+ robots with a total sum insured above ¥200M, and they price on hardware value because they have no failure-frequency data verified.

What third-party maintenance teaches

Third-party maintenance means an independent firm services equipment made by someone else. It has worked in three mature industries, and each one tells you something different about robots.

The scale can be large once it works. Park Place Technologies and Service Express, the two big US IT maintenance firms, merged in Jan 2026 with Warburg Pincus as majority owner and Temasek as a significant minority, 3,000+ staff and 25,000+ customers verified; a reported US$2.8B debt package at about 7 times leverage implies roughly US$400M of annual earnings estimate. In China, JD.com is building the first large multi-brand robot repair network. Its engineers report misassembled parts that leave robots unable to stand verified Aug 2026.

Terms in this chapter