Why a Robot With 99 Percent Per-Action Success Finishes Only 37 Percent of Jobs

A robot that succeeds on 99 percent of its individual actions completes a 100-action workflow on the first attempt about 36.6 percent of the time. That figure comes from an Oct. 10 essay in The Robot Report by Nicolas Sauvage, founder and president of TDK Ventures, and it is the cleanest way to read any humanoid demo.

Sauvage launched TDK Ventures, the corporate venture arm of TDK Corp., in 2019, and the firm manages $500 million across four funds. He invests in the category he is describing, which is worth holding next to his argument. The arithmetic below does not depend on him, though: it is a product of independent probabilities, and anyone can reproduce it.

The compounding arithmetic

If every action succeeds with probability p and the actions are independent, a workflow of n actions succeeds with probability p raised to the n. Sauvage's three examples use n = 100.

1. 99% per action36.6 %
2. 99.9% per action90.5 %
3. 99.99% per action99 %

Source: Nicolas Sauvage, TDK Ventures, in The Robot Report, Oct. 10, 2026; assumes equal, independent per-action success

Each extra decimal place of per-action reliability buys back most of the missing success. Going from 99 percent to 99.9 percent lifts the workflow figure by 53.9 percentage points; the next step adds 8.5.

Sauvage's one-line version: "A robot with a 99% success rate is not necessarily 99% automated."

Two numbers we added

These are our own calculations from the same formula, not figures from the essay. A 20-action workflow at 99 percent per action succeeds about 81.8 percent of the time. To reach a 95 percent first-attempt rate over 100 actions, per-action success has to be about 99.95 percent.

The assumption of independence cuts both ways. Real failures cluster, since a slipping grasp makes the next action more likely to fail, which makes the true figure for some workflows lower and, where a robot can retry, higher.

What Sauvage says to measure instead

He argues dexterity is a closed loop across vision, position, contact, force, pressure and recovery, and that touch is not needed in every workflow but can detect slip and unstable contact. His proposed metrics are the share of complete workflows finished without human help, unassisted recovery rates, and performance under changes in objects, lighting, position and wear. He adds failure points under repeated use, repair time, uptime and cost per completed job.

He also argues against maximal hands. His examples of "minimum sufficient dexterity" are Agility Robotics' Digit, ANYbotics' four-legged inspection robots, which one deployment used for more than 33,000 inspections across 450 inspection points, and Starship Technologies' wheeled delivery robots, with more than 10 million autonomous deliveries reported. Those deployment counts come from the essay and are not independently confirmed here.

Sauvage adds a point that matters for anyone redesigning hardware mid-program. Deployment reveals slips, poor grasps, wear and failed recoveries, and that data improves software, simulation, sensing, control and hardware. But changing the hand changes movement, sensing and calibration, so data collected on one design may not transfer to the next. A vendor that swaps its end effector therefore resets part of its own learning curve.

How Boston Dynamics' new hand fits

Boston Dynamics unveiled a 13-degree-of-freedom Atlas hand on Oct. 1, and IEEE Spectrum reported that Alberto Rodriguez, director of robot behavior for Atlas, called hands "a ruthless design trade-off." Rodriguez told The Robot Report on Oct. 9 that "a reliable robot allows you to capture data faster, run more experiments, and take bets at early deployments."

That is Sauvage's argument from the hardware side: reliability sets how fast a fleet learns. Boston Dynamics has published no per-action success rate, cycle count or workflow completion figure for the hand, so the chart above cannot yet be filled in for Atlas.

For how data collection feeds this loop, see our report on Mecka AI's $60 million robot-data round and on SafeWorld's $12.2 million seed for robot safety simulation.

A buyer reading a vendor deck can apply the formula in one step: take the per-action success figure on the slide, raise it to the number of steps in the buyer's own job, and compare the result with the completion rate the vendor quotes. If the deck gives only the first number, the second is unknown, and that gap is the question to ask.

The next place a vendor can be asked for a workflow completion rate is RoboBusiness 2026, Oct. 20 and 21 in Santa Clara, Calif.