Physical AI Decision Lab
Software that makes a bad call can be rolled back. A robot that makes one has already moved. This lab is for the person who signs off the deployment: run the floor, question the vendor, and decide what evidence would let you say yes.
Vendors quote success per action. Your operation experiences success per job. A pick is a dozen small actions in a row, and every one must go right. Move the dials and watch where the human hours go.
Detect, approach, align, grasp, lift, verify, place: each one counts.
| Per-action success | Jobs without help | Failed jobs per robot-hour | Robots per person |
|---|
Arithmetic, not a forecast: it assumes failures are independent and every failed job needs a person. Real failures cluster around particular items, shelves and shifts, which is why the next station runs a floor instead of a formula.
One robot, a set of pickers and a warehouse with racks that block line of sight. Choose how the robot behaves, how fast it may go and how good its sensors are. Then run 40 shift segments and see what you bought: throughput, or near misses.
Share of sensor readings that miss a visible person: glare, dust, a dirty lens.
Every setting you run is plotted below, so you can see the trade-off you are choosing rather than a single number.
Each dot is one setting, run over the same 40 segments with the same random seeds, so differences come from your choices. Lower and further right is better. Risk events = contacts × 10 + near misses, per robot-hour.
The floor is a 26 × 13 grid of one-metre cells; one step is one second. Pickers walk to random shelf faces, pause to pick and move on, and they step aside for a robot they can see. The robot fetches from a random shelf and returns to the dock. Its sensors see a person only within range, only along a clear line of sight past the racks, and miss a share of observations at the rate you set; a tracked person is remembered for two seconds. Braking is limited to 1 m/s², so stopping distance grows with speed. A contact is the robot entering a cell a person occupies, usually because someone stepped out inside its stopping distance; each costs 20 seconds of downtime. A near miss is a person found directly in front and unseen, or passed within a metre at 2 m/s or more. Each segment is 15 simulated minutes; 40 segments make 10 robot-hours. It is a working model that runs in your browser, not a validated safety case for any product.
You run operations for a four-site warehouse network. A robotics vendor wants Site 2. The board meets on Friday, and you have time to read three reports before then. Which would you check first?
“Proven at a reference customer. Certified safe around people. Pays for itself in just over a year.”
Take this into the meeting. For each question, mark what you have actually seen. A claim is not evidence, and evidence from someone else’s floor is not evidence from yours.
Rate every claim by the rung it stands on. A launch video is rung one, however good it looks. A sound deployment decision rests on rungs four and five.
A video or live demo. Shows the robot can do the task at least once, in conditions someone chose.
Success rates from the vendor’s own sites. Ask how success was defined, whether retries counted and who was standing nearby.
Fixed criteria, held-out environments, failures counted strictly. Figure’s Helix 2.5 reported 56% full-task success in 30 homes it had never seen, with safety interventions counted as failures. Read the analysis.
Intervention logs, availability and incident records from a customer like you, obtained from that customer rather than from the deck.
A bounded pilot in your own layout and traffic, with a comparison area, stop conditions and an incident log you own.
Each sounds reasonable in a boardroom and fails on a warehouse floor.
Capability and assurance are separate engineering problems. A smarter model can be confidently wrong in a situation nobody tested. Safety comes from limits, stops and supervision that still work when the model does not.
A stop is the system knowing when not to act. Measure stops, explain them and reduce them, but never tune them away to hit a throughput target. On bounded autonomy.
Pilots run with vendor engineers on site, a tidy zone and an attentive team. Scale runs on your night shift. Write down which pilot conditions will not exist at scale, then test without them.
The shape should follow the task. In structured spaces, wheels, conveyors and fixed arms often do the job with fewer ways to fail. A human shape earns its cost where the environment was built for people and cannot change.
Responsibility is moving explicitly onto the organisation that runs the robots, not only the one that builds them.
The revised international safety standard for industrial robots, the baseline robot makers design against.
The US mobile-robot standard’s part for users: risk assessment, management of change, training and periodic review in day-to-day operation.
Replaces the Machinery Directive and addresses machinery whose safety functions rely on AI that learns.
For deployments in India, confirm the requirements that apply to your site with your own safety and legal advisers. Nothing here is legal or regulatory advice.