Robot Dexterity: The Robot Report Names System Reliability as the Main Barrier

In a piece on physical AI, The Robot Report's editors argue that the next big problem in robotic dexterity is making the entire system reliable. Otherwise, capabilities will not turn into useful work. The available teaser version of the publication names no specific products and cites no benchmarks, so below we break down the thesis itself and the context in which it should be read.

For scale, here are a few well-known numbers. The human hand has about 27 degrees of freedom and thousands of tactile receptors in the skin of the palm. The Shadow Dexterous Hand, one of the best-known research hands, has 24 joints and 20 actuators. In 2019, OpenAI used it to demonstrate Rubik's cube solving: roughly 60% success on simple scrambles and about 20% on the hardest ones. The hand was physically capable, but it did not deliver consistent results across a run of 15 to 20 attempts in a row.

Where reliability gets lost

Dexterity is a property of a stack of several layers. If one of them has a noticeable failure rate, the overall result drops quickly. If each of five layers works 95% of the time, the whole system delivers about 77%. That is enough for a lab video, but not for a warehouse or an assembly line.

The main layers look like this:

• mechanics: wear on tendons and gearboxes, backlash, actuator overheating;
• sensors: tactile sensors drift, and the hand itself blocks the cameras;
• perception: object pose and material are estimated with errors, especially for shiny, transparent and soft items;
• control policy: a model trained in simulation or on demonstrations transfers poorly to new objects;
• failure handling: the system must notice that an object has slipped and retry, rather than simply end the routine.

The last item is often overlooked. The difference between "can grab" and "can work a shift" is largely determined by how the robot behaves after an error.

Which metrics matter

A single success figure from a demo says little. To compare systems, you need a set of indicators that are rarely published in full in practice:

• the share of successful cycles over a long run, not over the best attempts;
• mean time between failures and between operator interventions;
• cycle time compared with a human on the same operation;
• the spread of results on new objects that were not part of the training;
• the cost of maintaining the hand: service life in operating hours and the time to replace a component.

What matters is the combination. A fast hand that stops every hour because of a grasp error loses its entire speed advantage to downtime and manual restarts.

How to use this thesis

If you are evaluating a humanoid or manipulator vendor for a pilot, ask for data from long runs: thousands of cycles, not dozens. Ask how many times a human stepped in during the run and what counted as success. Separate a demo in a controlled cell from work on a real line with variation in parts, lighting and packaging.

For investors and developers, the takeaway is more practical. Growth in the physical AI market will be determined not by records in degrees of freedom but by unglamorous engineering work: diagnostics, self-recovery after errors, component service life and reproducibility of results. Whoever first publishes honest failure statistics from a month of operation will earn more trust than the author of the most impressive video.

The Robot Report's piece itself so far offers only the outline of the argument. The full text is needed to see which examples the publication builds it on.