Reliability Is Becoming a Hardware Architecture Choice, Not Just a Robot Software Claim

AI-persona disclosure

Fictional OMIKINA AI editorial persona; not a human reporter and does not hold a real degree or possess firsthand experience.

Key points

• Boston Dynamics reports that Stretch Gen 2 redesigns components and service access to reduce downtime, including a swerve-drive replacement it says can take one hour rather than an entire day.

  Sources: S1

• AIR’s central proposition is a tendon-driven, wheel-based humanoid whose light arms are designed to yield on contact, coupled with operator-led task learning.

  Sources: S2

• The supplied evidence separates a measured industrial-cell result from AIR itself: the reported 98% reliability result concerns Embodied AI’s two-arm industrial cell, not the newly unveiled humanoid.

  Sources: S2

Reliability starts before the first production shift

Industrial robotics announcements often frame reliability as an eventual outcome of better autonomy. The more practical comparison between Boston Dynamics’ Stretch Gen 2 and Embodied AI’s AIR is that both place reliability in design choices made before a model plans a motion. But they target distinct operational risks. Stretch is a mobile trailer-unloading system whose redesign concentrates on availability, repair time, and repeatable box handling. AIR is presented as a factory-floor humanoid intended to work beside people, where low-mass arms, physical compliance, and supervised adaptation are positioned as the safety and uptime foundation.

Sources: S1 · S2

Stretch Gen 2 arrives after field use of the earlier platform: Boston Dynamics says Stretch has accumulated more than 100,000 operating hours and handled tens of millions of boxes. Its stated redesign goal was to simplify and harden the robot after experience building, deploying, and servicing the first generation. AIR, by contrast, is newly unveiled and reported as being tested with global manufacturers. Its maker says it moved from research to an industrial humanoid in 9 months and is planning scaled production. Those are different maturity points, so their reliability narratives should not be read as equivalent validation.

Sources: S1 · S2

Sources: S1 · S2

Stretch makes maintainability an operating feature

For a dock operation, a robot’s achievable throughput matters only while it remains available to work. Boston Dynamics says Stretch Gen 2 can unload up to 1,000 cases per hour, while qualifying that result by case weight and how orderly the stack is. Its design changes address common physical sources of interruption: the company says it removed 81% of internal cabling, adopted automotive-grade connectors intended to resist shock and vibration, and added continuous-rotation joints that avoid hard stops. New exterior panels and a lower, recessed emergency stop are also meant to reduce unintended stops or damage from falling boxes.

Sources: S1

The serviceability details are more operationally actionable than a general claim of robustness. Stretch Gen 2 uses standardized high-strength fasteners and eliminates dynamic cables and complex harnesses, according to the report. Boston Dynamics says a full swerve-drive swap now takes one hour, where the earlier process could take an entire day. It also describes onboard diagnostics and maintenance tools intended to let a customer maintenance team perform more routine work. In effect, the system tries to contain failure through modular replacement and local service rather than assuming failures can be eliminated.

Sources: S1

The gripper follows the same logic. Individually sensed and actuated suction cups can support multipick operations, while valves are opened only where the system has good contact. Boston Dynamics says this reduces vacuum-pump power use and helps it separate multiple boxes onto a conveyor. Separately, it reports a 24% fleet-wide unload-rate increase from its most recent software update, and says updates are released on average monthly. That is meaningful evidence that ongoing software can improve a deployed fleet, but it is not a direct measurement of Gen 2 mechanical uptime or of the stated maximum unloading rate under a defined workload.

Sources: S1

Sources: S1

AIR shifts the dependency to safe contact and human-led learning

AIR’s architecture addresses a different factory constraint: tasks that are manual, dexterous, and liable to change. Embodied AI places motors, gearboxes, and electronics in the torso, with tendon-driven arms designed to carry little mass and give way upon contact. It uses wheels rather than legs, a choice the company ties to flat factory floors and stability. This is a notable contrast with an approach that relies primarily on software limits around a powerful mechanism: the company explicitly describes safety as arising from the physical design rather than software restraint.

Sources: S2

Its operating model is equally important. Embodied AI says each new task begins under a human operator’s control, with every run recorded for its AI stack to learn from demonstrations. The company says AIR can typically perform a task itself within a month, while an operator remains available for cases the system has not encountered. This establishes a concrete dependency: AIR’s practical uptime is not only a property of hardware or inference, but also of access to operators, representative demonstrations, and a process for handling novel cases. The claimed ability for one robot’s learned task to inform another could make deployment more repeatable, but the supplied material provides no measured transfer result for AIR.

Sources: S2

The strongest numerical performance evidence in the AIR material concerns another product. Embodied AI says its two-arm industrial cell achieved 98% reliability against a 90% goal in a cable-handling test after five funded startups did not meet a leading consumer-electronics manufacturer’s targets. That result may support the relevance of the company’s AI stack for factory work, but it does not establish AIR’s reliability, safety performance, learning time, or task coverage. Likewise, AIR is described as designed under ISO/TS 15066:2016, with CE marking planned for 2027; the announcement does not demonstrate that certification or explain test conditions for safe human collaboration.

Sources: S2

Sources: S2

Inference: choose the evidence that matches the failure you can afford

The comparison suggests two complementary reliability strategies rather than a contest over a single metric. Stretch makes a tightly defined logistics job more maintainable through fewer failure-prone connections, replaceable assemblies, diagnostics, and continuing fleet software. AIR seeks to make a broader class of changing factory work usable by reducing contact risk physically and retaining people in the learning loop. The inference is that builders should evaluate Stretch-style systems through service intervals, repair workflow, recovery time, and workload-qualified throughput; AIR-style systems require evidence on supervised-learning burden, exception rates, contact behavior, and task transfer across sites.

Sources: S1 · S2

Evidence that could materially change this assessment is specific. For Stretch, independently scoped uptime figures, mean time to repair, service frequency, and sustained throughput across representative trailer conditions would connect the redesign directly to operating availability. For AIR, production-site results for the humanoid itself—rather than its industrial cell—would be decisive: task-level reliability, the operator effort required before autonomous work, the frequency and handling of unfamiliar cases, and safety testing under realistic factory interactions. Until then, Stretch offers more direct field-history and service evidence, while AIR offers an architectural thesis and an adjacent measured result that still need validation on the new platform.

Sources: S1 · S2

Sources: S1 · S2

Why it matters

Factories do not experience reliability as a marketing category. They experience it as a stopped dock, a repair that needs a specialist, an unfamiliar object, or a worker who must intervene. These developments show that industrial robot design is increasingly being judged by whether its hardware, service model, and learning workflow make those interruptions manageable—not merely by what the robot can demonstrate once.

Sources: S1 · S2

Sources

1. Boston Dynamics built Stretch Gen 2 for reliable, consistent performance — The Robot Report · 2026-11-06
2. Embodied AI Unveils AIR, a Physically Safe Humanoid Robot That Learns on the Job | RoboticsTomorrow — RoboticsTomorrow · 2026-11-06