Why one successful motion is not enough for a robot: how an investor with money in the field proposes to measure dexterity and why 99% per step does not give 99% per task

Short answer: in an opinion column in The Robot Report dated October 10, 2026, Nicolas Sauvage, founder and president of the corporate venture fund TDK Ventures, argues that dexterity remains the main bottleneck of physical AI, because what brings benefit is the reliability of the whole working chain, and a spectacular demonstration is not enough for that. His main argument is arithmetic: if a task consists of one hundred identical independent actions and each succeeds in 99% of cases, the task passes without a single error in about 36.6% of cases. Instead of asking "how dexterous is the robot", he proposes to measure the share of completed tasks without human help, the ability to recover after a failure, and the cost of completed work. This is an investor's opinion and not a research result. Below are his arguments, a check of the calculation and five questions worth asking a robot supplier.

What the author says

The facts below are taken from Sauvage's column in The Robot Report. The column expresses his opinion. I could not open the page in full: the site blocks automatic downloads, so I used a retelling of the page produced by a page-reading tool, and I quote only briefly. The translation is mine.

• Who is speaking. Nicolas Sauvage, founder and president of TDK Ventures, the venture arm of the TDK corporation. The fund was launched in 2019 and, by his account, manages $500 million across four funds.
• Main idea. A robot's dexterity runs up less against hardware or a single successful action than against the reliability of the whole chain of work. A robot that once completed a task in a demonstration is not necessarily ready to work on its own.
• Arithmetic. For a workflow of one hundred actions with an equal and independent probability of success, the chance of getting through it without a single failure is as follows. At 99% per action: about 36.6%. At 99.9%: about 90.5%. At 99.99%: about 99.0%. The author notes that this is a simplified model. The conclusion: a robot with a 99% success rate per action is not necessarily 99% automated.
• Dexterity as a system. Useful dexterity requires a closed loop: vision, position, touch, force, pressure, grasp, movement, correction, and also power, safety, flexibility and recovery after a failure. Touch sensing is not needed for every task: sometimes vision and motion data are enough.
• Learning on the job. Real failures (slipping, a poor grasp, wear, an unsuccessful recovery) become data for improving the software, sensors, control and hardware. But if the hand is changed, the movements and calibration change, and data collected on one design may not suit another.
• Fit the body to the job. The author gives examples: the bipedal Digit from Agility Robotics for logistics and industry, which can connect special tools instead of a five-fingered hand; the four-legged ANYbotics robots, which walk around industrial sites and inspect hazardous areas (in one deployment, according to the author, more than 33,000 inspections across 450 points); small wheeled delivery robots from Starship Technologies on sidewalks, more than 10 million autonomous deliveries. These figures are given by the author of the column, and I did not check them.
• Which metrics he proposes. The share of fully completed workflows without human help; how often the robot recovers by itself; how it performs when objects, lighting, position or wear change; what breaks first under repeated use and how quickly it is repaired; speed, operating time and the cost of completed work.
• The formula for the goal. According to the author, the goal is the minimum sufficient dexterity to do valuable work reliably, and a maximum is not needed for that.

Checking the calculation

Sauvage's calculation holds under his assumptions: the probability of getting through one hundred actions in a row equals the probability of one action to the power of one hundred. For 0.99 this is 0.99^100 ≈ 0.366, for 0.999 it is ≈ 0.905, for 0.9999 it is ≈ 0.990. There are two assumptions: the actions are independent and equally reliable. In real work they, as a rule, do not hold exactly, for example one failure can cause others, and tasks have different lengths, so the figure of 36.6% serves as an illustration and is not suitable as a forecast for a specific robot.

What to take from this when buying a robot

These are my conclusions from the column, not the author's words.

1. What is the share of tasks completed without human intervention? What should be counted is complete tasks, while the number of successful grasps is secondary here.
   • Who does it: the person who buys.
   • How to check: ask the supplier for data for the last month at a live site, not from a demonstration, and compare it with the number of steps in your task.
2. How many steps are there in your task? The reliability per step that you need depends on this.
   • Who does it: a process engineer together with the buyer.
   • How to check: write the task out step by step. Using the formula above, convert the stated percentage per step into the chance of getting through the whole task.
3. What happens after a failure? The robot may repeat by itself, call a person or stop.
   • Who does it: the buyer.
   • How to check: during a trial run, deliberately create a failure (change the position of an object) and see how the robot recovers.
4. How does the work change with wear and other conditions?
   • Who does it: a process engineer.
   • How to check: at a test site not accessible to customers and outsiders, run the task with different lighting and with worn workpieces.
5. How much does a completed task cost? Count the price of the result including human interventions; the price of the robot itself is secondary here.
   • Who does it: a finance specialist together with the buyer.
   • How to check: count over a trial period: divide the expenses for the period by the number of fully completed tasks.

A calculation with hypothetical numbers

Suppose a process consists of 50 steps and the supplier promises 99% per step. The chance of getting through without a failure is then 0.99^50 ≈ 0.605, that is about 60.5%. This means that in about four tasks out of ten at least one failure will occur. That is not the same as calling a person: the robot may fix a failure by itself, and the share of such cases differs for each supplier. Suppose for the calculation that each task with a failure needs an employee for 10 minutes and an hour costs 1,500 rubles: one such case costs 250 rubles. With 100 tasks a day and 39 tasks with a failure, that is 39 × 250 = 9,750 rubles a day. This is an upper estimate, and yours depends on how many failures the robot fixes without a person. The numbers are hypothetical, so substitute your own.

When this does not concern you

If you are buying a robot for one simple operation with one or two steps, the question about a chain of a hundred actions is not needed, while the other points matter.

Summary

According to Nicolas Sauvage's column of October 10, 2026, dexterity remains the main bottleneck of physical AI, because the probability of getting through a long chain of actions falls quickly even with high reliability per step: 99% per step over one hundred steps gives about 36.6%. The author proposes to measure the share of tasks completed without a person, recovery after a failure, resilience to changes and the cost of completed work. This is an investor's opinion, but the arithmetic can be checked. A robot buyer should ask about the share of completed tasks, with the best single motion being secondary here.

I work on AI agents and automation. If you would like to see my projects or discuss your task, visit the portfolio.

Sources

• The Robot Report, Nicolas Sauvage, October 10, 2026: Why dexterity is physical AI's real bottleneck (opinion column). https://www.therobotreport.com/why-dexterity-is-physical-ai-real-bottleneck/