Physical AI’s reliability test is no longer just task success—it is recovery
A factory humanoid, warehouse inventory drones and a laboratory research system point to the same operational lesson: useful physical AI needs bounded work, independent checks and a way to resume after the world changes.
By Nia Okafor · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human security credentials or firsthand experience.
Key points
- Google DeepMind reports that its Gemini Robotics 2.5 pilot with Apptronik’s Apollo 2 achieved a stated success rate in automotive kitting, while its broader evaluations measured task performance across a large set of manipulation tasks.
Sources: S1
- ASICS says Corvus inventory drones made discrepancies visible sooner by increasing count frequency in selected warehouse areas and shifting manual effort toward investigation and resolution.
Sources: S2
- Robo-MLT’s laboratory results suggest that explicit monitoring and replanning can improve recovery when a workflow is interrupted, though the evidence is limited to the reported evaluation settings.
Sources: S3
The important reliability question is what happens after a miss
Physical AI is moving beyond demonstrations in which a robot completes an isolated motion. The more consequential question for an operator is whether the system can detect that its understanding is wrong, contain the impact and restore useful work without turning a small exception into a stalled process. The evidence across industrial kitting, warehouse inventory control and laboratory manipulation points toward a practical distinction: reliable automation is not necessarily automation that never encounters an error. It is automation designed around a narrow operational role, a visible source of truth and a defined recovery path.
Google DeepMind’s reported Mercedes-Benz proof of concept is a useful example of a bounded role. Apollo 2 sorts metal automotive brackets from bulk bins into sequencing containers in the order needed for production, with Gemini Robotics ER tracking the workflow and Gemini Robotics 2.5 controlling hand and arm actions. DeepMind reports a 96% success rate for this automotive-kitting pilot. That is meaningful evidence that a humanoid can contribute to a repetitive physical workflow, but it is not evidence of general reliability across a factory. The stated result belongs to a particular task, robot, partner sites and proof-of-concept conditions.
Sources: S1
The remaining failures matter more than the headline percentage may imply. In a production setting, the operational cost of an incorrect pick, an unfilled container or a robot that fails to recognize a changed condition depends on whether the line can identify and correct it before the error propagates. DeepMind says its model uses inference-time reasoning guidance to steer around hazards and reports improvements in safe task completion around obstacles and nearby people in its evaluations. Those are prevention-oriented controls. The supplied account does not detail the recovery procedure for a failed automotive-kitting attempt, including when the robot stops, how an exception is handed over or how a mistaken placement is verified.
Sources: S1
Inventory drones show why detection can be a product feature
The ASICS deployment offers a less glamorous but more operationally mature model of resilience. Corvus One flies nightly in two high-risk storage areas at the company’s Byhalia distribution center, reading UPCs and counting boxes. ASICS reports that those areas moved from being counted twice yearly to monthly, and that misplaced inventory is recovered roughly 60% faster. The system’s value is not simply autonomous flight: it shortens the period in which a warehouse record and the physical inventory can disagree.
Sources: S2
That design includes a concrete cross-check. In the singles area, a UPC scan alone can be associated with the wrong box because returns can put the same SKU in a location more than once. Corvus therefore combines box counts with UPC reads; ASICS describes comparing the physical count with the number of reads. The following morning, an inventory associate reviews each flight and resolves most discrepancies from a desk using captured video and images. Here, autonomy does not remove people from the exception loop. It moves their work from routine aisle counting to evidence-based reconciliation, root-cause research and other exception work.
Sources: S2
ASICS reports redirecting about 200 hours of monthly counting labor and finding misplaced cases during a reported three-month period that otherwise might have remained unavailable until the next physical count. It also says it now publishes its full marketplace availability feed rather than withholding inventory because of fulfillment risk. These are company-reported operating outcomes, not a controlled comparison of every warehouse condition. Still, they show a critical dependency often obscured in physical-AI claims: a robot’s sensing output only improves decisions when it is connected to a review workflow and an inventory system that can act on discovered discrepancies.
Sources: S2
Sources: S2
Monitoring changes recovery from an afterthought into a system function
The Robo-MLT research makes the recovery mechanism explicit. Its framework separates learned action generation from workflow monitoring: a vision-language planner creates an ordered skill sequence, while a multiframe monitor determines whether the active skill has completed or become infeasible and either advances the workflow or triggers replanning. In the reported blood-gas-analysis and blood-sample-tube-sorting evaluations, Robo-MLT achieved task success rates of 82% and 70%, respectively, and exceeded the cited π0.5 baseline by 30 and 22 percentage points.
Sources: S3
More revealing than the nominal scores is the reported interruption case. When an operator removes the sample currently being handled, Robo-MLT revises the remaining plan and continues with other samples. This is a different reliability claim from dexterity or obstacle avoidance. The system is not merely trying to prevent an unsafe or unsuccessful movement; it is representing the workflow state sufficiently to recognize that its original plan no longer applies. The supplied abstract says the study evaluated nominal configurations and sample-removal configurations, so its conclusions should remain confined to those laboratory tasks and tested disruptions.
Sources: S3
Inference: taken together, these developments suggest that the nearer-term winning pattern for physical AI is supervised autonomy with layered controls, rather than a promise of universal robot competence. Gemini Robotics emphasizes broad manipulation capability and safety steerability; Corvus operationalizes frequent sensing plus human exception review; Robo-MLT formalizes state monitoring and replanning. They solve different problems and should not be treated as interchangeable benchmarks. But each reinforces the idea that a useful deployment needs a reliable way to compare the system’s belief with the physical world, then decide whether to continue, escalate or replan.
What operators should ask next
The available evidence supports cautious optimism, not a general reliability guarantee. DeepMind reports evaluations across more than 200 tasks and describes simulated practice runs used to help robots generalize, while the factory result remains a specific kitting pilot. ASICS describes an installed warehouse system with recurring flights, but its performance figures are reported by the deployment participants. Robo-MLT reports comparative results and recovery behavior, but the supplied material is an abstract and identifies limited laboratory workflows. None of these accounts establishes how performance holds under every equipment fault, ambiguous object state, communications problem, shift change or prolonged sequence of exceptions.
The assessment would materially change with evidence that connects task success to operational recovery: rates of detected versus undetected errors, the time and human effort needed to resolve each exception, behavior after repeated disruptions, and results across changed layouts, objects and procedures. For systems near people, operators also need evidence that hazard avoidance, stop behavior and resumption are tested as one workflow rather than as separate claims. The practical procurement question is therefore not only whether a robot can do the task. It is whether the surrounding system can prove when it did not, limit the consequence and get work moving again.
Why it matters
Physical AI can create value before it becomes broadly dependable, but only where operators engineer for mismatch between digital plans and physical reality. The strongest near-term deployments pair automation with verification, accountable exception handling and a documented route back to productive work after prevention fails.
Sources
- Gemini Robotics 2.5 accelerates real-world general dexterity — Google DeepMind ·
- ASICS cuts inventory recovery time 60 percent using Corvus Robotics drones — Robotics & Automation News ·
- Robo-MLT: Dual-System Robotic Manipulation for Long-Horizon Medical Laboratory Tasks — arXiv Robotics ·