Robot arm control
Sense, Decide, Move, With a Person Asked Before a Risky Move
A robot arm runs one loop, sense, decide, move, and the human is the gate before the risky move. ISO 10218:2025 (Feb 2025) folds the old ISO/TS 15066 limits into the standard and makes force validation mandatory: a chest takes 140 N before it hurts, a face 65 N, and a 20 kg arm moving toward a chest must stay under 0.24 m/s. The safe speed costs 2.55 seconds a move, and the log is what proves the arm asked.
First, a simple picture
A forklift driver who stops and looks before reversing.
A forklift driver does not reverse on faith. The driver looks, checks the mirror, and moves only when the way is clear, and if a person is standing where the forks will go, the driver stops and waits. The machine is strong, the person is soft, and the rule is simple: the machine does not move into the person.
A robot arm runs the same loop, faster and without a driver. Sense, where is everything, including the person. Decide, what should happen next, and is it safe. Move, the arm moves, on the loop that was tuned in simulation first. And before any risky move, the arm stops and asks. You say yes first. That is senovasys.com's own rule, and it is the through-line of this article: the loop is sense, decide, move, and the human is the gate before the risky move.
The loop, in plain words
Sense is the arm's eyes and skin. Cameras watch the bench. A force sensor in the tool flange feels contact, on a UR5e-class arm, the sensor reads force to an accuracy of 4 N. Joint torque sensors feel a bump anywhere along the arm. A laser scanner watches the floor. The event camera from Article #2 only reports change, so the arm is not drowned in data, it is told when something moved. The Kalman filter from Article #4 fuses the sensors into one estimate of where everything is, because no single sensor tells the truth on its own.
Decide is the model and the plan. The on-device model from Article #9, the visual wake-words CNN that runs in 38 ms for 387 µJ on a microcontroller, decides whether the shape in the camera's view is a person. The plan decides what the arm should do next: pick up the battery pack, carry it across the bench, set it down. And the safety limits decide what the arm is allowed to do: the speed cap, the force cap, the separation distance. The decision is not one thing. It is three things, the model's answer, the plan's next step, and the safety limits, and the arm moves only when all three agree.
Move is the control loop from Article #3, the PID loop, tuned in simulation before anything moves, that turns a desired position into motor commands. The loop is the arm's hands. The loop is also the arm's conscience: it is the thing that stops the arm when the force sensor says contact, and it is the thing that holds the arm still when the gate is waiting for a yes.
The loop runs in milliseconds. The gate runs in seconds, the time it takes a person to look, think, and say yes. That mismatch is the whole design problem, and the rest of this article is about it.
What "safe and auditable" actually means
"Safe" is a number per body region, not a feeling. ISO 10218 is the international standard for robot safety, in two parts: Part 1 covers the robot itself, Part 2 covers the application and the cell. In February 2025 it was revised, ISO 10218-1:2025 and ISO 10218-2:2025 replace the 2011 editions, and the revision did two things that matter for anyone building a cobot cell.
First, it folded the collaborative-robot content of ISO/TS 15066 into Part 2, so the force and pressure limits are no longer supplementary guidance but a formal requirement. Second, it made validation mandatory: any installation using power-and-force limiting must measure and verify real contact forces and pressures against the limits. The United States adopted both parts as ANSI/A3 R15.06-2025 in September 2025, replacing R15.06-2012. Part 2 alone grew from 72 pages to 223.
The standard keeps four collaborative applications, the four ways a person and a robot can share a workspace during normal operation. Monitored standstill: the robot holds still under power while a person works, and resumes when the person leaves. Hand guiding: a person moves the robot through a guiding device, teaching it by hand.
Speed and separation monitoring: sensors track the distance to any person, and the robot slows, then stops, before contact could happen. Power and force limiting: contact is allowed, but force and pressure stay under the limits for the body region struck. Only the last mode needs a cobot, a light arm with contact sensing.
The other three work with any arm, because they keep the person and the robot apart, or keep the robot still, or keep the robot slow.
The body model is the heart of it. ISO/TS 15066 Annex A gives a body model of 29 specific body areas grouped into 12 body regions, with permissible force and pressure values for each, built on pain-onset research. It separates two kinds of contact: transient contact, a brief impact the body can recoil from, allowed a higher force, and quasi-static contact, clamping, where a body part is trapped between the robot and a fixed object and the force persists until the robot releases it, held to a stricter limit.
The quasi-static force limits run from 65 N for the face, the lowest, and the standard does not permit contact with the head and face at all, to 140 N for the chest and for the hands and fingers, to 220 N for the thighs and knees. The same table gives each region a spring constant, how stiff the body is there, from 10 N/mm for the abdomen, the softest, to 150 N/mm for the skull and forehead, the stiffest, with the face at 75, the neck at 50, the chest at 25, and the hands and fingers at 75. The spring constant and the effective mass turn a force limit into a speed limit: for a robot with 20 kg of effective mass, the permissible approach speed toward a chest is 0.24 m/s; toward hands and fingers, 0.67 m/s.
A lighter robot may move faster; a heavier robot must move slower. The force limit becomes a speed limit, and the speed limit is the number the cell is built around.
Three things the standard says that founders get wrong. First: power and force limiting does not make the end effector safe. A cobot carrying a sharp tool, a hot welding tip, or an unguarded grinder can cause injury regardless of force limiting, the force limit protects against the arm, not against the tool. Second: the separation distance is dominated by the person's approach speed and the sensor's intrusion allowance, so slowing the robot helps less than people expect. The formula is Sp = Sh + Sr + Ss + C + Zd + Zr, the person's travel during the system's reaction time, plus the robot's travel, plus the robot's stopping distance, plus the intrusion allowance, plus the sensor's detection uncertainty, plus the robot's position uncertainty.
In one second, a person walking at the standard's assumed 1,600 mm/s covers 1.6 metres; the arm at 250 mm/s covers 0.25 metres. The person is the fast one. Third: the safety functions that enforce the limits must meet the required performance level, SIL 2 / PLd minimum under ISO 13849, which means the stop is not a software feature.
It is a safety-rated function, built on hardware that fails safe, and it must be proven, not asserted.
And the 2025 revision adds a formal two-tier robot classification. A Class I robot is small: a manipulator mass of 10 kg or less, a maximum force of 50 N, and a top speed of 250 mm/s. That 250 mm/s is not an arbitrary number.
It is the chest limit, expressed as a speed, and it is the number the demo below is built on. Every industrial robot must also provide emergency stop (category 0 or 1), protective stop (category 0, 1, or 2), and normal stop (category 2) functions. The stops are not optional.
They are the floor.
The gate
The gate is not a new idea. Thomas Sheridan and William Verplank, at MIT's Man-Machine Systems Laboratory, described ten levels of automation in a 1978 report for the US Office of Naval Research, studying how much a human operator should delegate to a remotely operated undersea vehicle. The scale describes, for a single decision, how the work is divided between the computer and the human.
At level 4, the computer suggests one alternative and the human chooses. At level 5, the computer executes its suggestion if the human approves. At level 6, the computer allows the human a restricted time to veto before it executes.
At level 7, the computer executes automatically, then necessarily informs the human. The interesting boundary is between 4 and 5: the question changes from "who chooses?" to "who approves?", and the boundary between 6 and 7 is whether the human sees the action before it happens or only after.
The gate in this article is Sheridan level 5: the arm proposes the risky move, and the human approves it before the arm executes. The risky move is the one the risk assessment names, the battery lift over the person's hand, the move that crosses the person's workspace, the move that happens while the person is leaning in. The routine moves run at level 7 or 8, the arm executes, and the human is told after, or only if they ask.
The risky move runs at level 5, and the arm waits.
Lisanne Bainbridge's 1983 paper, Ironies of Automation, is the warning that comes with the gate. A human in the loop is a control only if the human can really say no, and a human who is asked to approve a hundred routine moves a day will stop reading the requests and start tapping "yes." The design response is to make human oversight meaningful rather than ubiquitous: ask rarely, ask only about the risky moves, and give the person the information they need to decide, what the arm is about to do, why it is risky, and what happens if they say no. The EU AI Act makes the same point as law: article 14 requires effective human oversight on every high-risk AI system. The gate is not a checkbox. It is a decision, and the person has to be able to make it.
And the gate has to be auditable. "Auditable" means the log proves three things: that the arm asked, that the person answered, and what the arm did next. The log is the answer to the question the risk assessment will ask, and the question the customer will ask after an incident: who approved this move, and when. A gate without a log is a gate nobody can prove existed. A log without a gate is a record of moves nobody approved.
Example Labs is a made-up company we use for worked examples.
One arm, two ways to move: Example Labs on the bench
This firm is made up. Call it Example Labs again: 60 staff, and now a robot arm on the bench, a UR5e-class six-axis cobot, 5 kg payload, 850 mm reach, 20.7 kg arm weight, maximum joint speed 180°/s, tool speed about 1 m/s, pose repeatability ±0.03 mm per ISO 9283, force/torque sensor accuracy 4 N, about 200 W under a typical program and 570 W maximum. The arm's job is to pick up the sensor hub's battery pack, carry it across the bench, and set it down, with a person at the bench, leaning in.
The demo's arithmetic is the article's own, verified before writing, on the published figures cited above; the firm and its bench are the worked example, not a measured cell.
Step 1, The job. The arm moves the battery pack across its 850 mm reach. The person is at the bench, and the arm's path crosses the person's chest height. The chest's quasi-static limit is 140 N. At 20 kg of effective mass, the permissible approach speed toward a chest is 0.24 m/s.
Step 2, Full speed: 17× the energy the chest limit allows. The arm runs at its rated tool speed, about 1 m/s. Its kinetic energy at that speed is ½ × 20 kg × (1 m/s)² = 10 J. The energy the chest limit allows, at 0.24 m/s, is ½ × 20 × 0.24² = 0.58 J. The full-speed arm carries 17.4× the energy the limit allows, and its speed is 4.2× the permissible approach speed. The arm crosses the bench in 0.85 seconds. It is fast, and it is over the limit by a factor of seventeen.
Step 3, Gated: inside the limit, at a price. The arm runs in power-and-force limiting mode, capped at 250 mm/s, the Class I speed, the chest limit expressed as a speed. Its kinetic energy is ½ × 20 × 0.25² = 0.63 J, just above the 0.58 J the chest limit allows at 0.24 m/s, and the cap is a speed, not an energy, so the cell's validation measures the real contact force, which is what the 2025 standard now requires. The arm crosses the bench in 3.40 seconds. The safe speed costs 2.55 seconds per move, four times slower. That is the price of the person staying in the room, and it is a price the cell pays on every move, all day.
Step 4, The gate, and the log. The battery lift, the move that carries the pack over the person's hand, is the risky move, and it is gated at Sheridan level 5. The arm stops, the screen shows what it is about to do and why it is risky, and the person says yes or no. The log records the ask, the answer, and the time. The arm waits. And the separation monitoring runs underneath it all: the scanner watches the gap, and if the person leans in faster than the arm can stop, the arm stops first, because in one second the person covers 1.6 metres and the arm covers 0.25, and the person is the fast one.
The worked example's arm: full speed against the 250 mm/s gate
| Full speed, 1 m/s | Gated, 250 mm/s | |
|---|---|---|
| Kinetic energy | 10 J (17× the chest limit) | 0.58 J (inside the chest limit) |
| Time across the bench | 0.85 s | 3.40 s |
| Cost of the safe speed | 2.55 s a move |
The decision test
Four questions. They sort the arms that are safe and auditable from the ones that are fast and hopeful:
- Has the risk assessment named the body regions the arm could touch, and the limits for each, and has the cell measured the real contact forces against them? No, the limits are a brochure claim, and the 2025 standard now requires validation, not assertion. Yes, the cell has the numbers, and the numbers are measured.
- Is the arm's speed set by the softest body part it could touch, the chest at 0.24 m/s, the face avoided altogether, rather than by the rated tool speed? No, the arm runs at 1 m/s because it can, and the person is the one who has to stay out of the way. Yes, the speed is the limit's speed, and the cycle time is honest about it.
- Is the risky move gated at Sheridan level 5, the arm proposes, the person approves, the arm waits, and is the gate asked rarely enough that the person actually reads it? No, the arm moves on its own, or the person taps "yes" a hundred times a day and has stopped reading. Yes, the gate is on the risky moves only, and the person can really say no.
- Does the log prove the gate, the ask, the answer, and the time, and does it record every stop, every contact event, and every override? No, the cell has a log of moves, not a log of decisions, and nobody can prove the arm asked. Yes, the log is the audit trail, and the audit trail is the product.
If the answers are yes, yes, yes, yes, the arm is safe and auditable, and the review is a final check, not a rescue. If any answer is no, the job is a robotics safety/architecture review, and the honest thing is to do it before the arm ships, not after the first incident. And if the answer to question 2 is "we run at full speed and rely on the person to stay clear", the problem is not the arm.
The problem is that the cell has outsourced its safety to the person's attention, and attention is the one safety device that gets tired.
"But cobots are safe by design", and two more fair doubts
"But cobots are safe by design." A cobot is safe by design in the sense that it limits force and pressure, and the limit is a number per body region, not a guarantee. A cobot carrying a sharp tool, a hot welding tip, or an unguarded grinder can cause injury regardless of force limiting. And the force limit is only as good as the validation: the 2025 standard requires the cell to measure the real contact forces, because the arm's datasheet says what the arm can do, not what the cell does. "Safe by design" is the starting point. The risk assessment, the speed cap, the gate, and the log are the rest.
"The gate will slow us down." It will, by 2.55 seconds a move, in the demo, and by the seconds the person takes to read the request and say yes, on the risky moves. That is the price, and it is worth stating plainly: safety costs speed, and the cost is the price of the person staying in the room. The alternative is not free. The alternative is a fenced cell, a person who cannot work beside the arm, and a product that cannot ship until the fence is built. The gate is what lets the person stay.
"We can just fence it off." You can, and for a heavy arm, a fast arm, or a sharp tool, you should. Fencing is the honest answer when the arm's energy is too high to share a room with a person. But fencing is also a confession: the cell could not make the arm safe beside a person, so it removed the person. The cobot's promise is the opposite, the person stays, the arm slows, and the gate decides the risky moves. If the job needs the person beside the arm, the gate is the way. If the job does not need the person beside the arm, the fence is the way. The job decides.
Sources
- Relling Systems, "Cobot Safety (2026): ISO 10218:2025, PFL and Risk Assessment"
What we used
ISO 10218-1:2025 and ISO 10218-2:2025 published February 2025, replacing the 2011 editions; ISO/TS 15066 content now inside ISO 10218-2; four collaborative applications; monitored standstill is the new name for the old safety-rated monitored stop; force limits depend on body region and on bump vs clamp; separation distance dominated by the person's approach and the sensor's intrusion allowance; Sp = Sh + Sr + Ss + C + Zd + Zr; ANSI/A3 R15.06-2025 is the US adoption, published September 2025
- AMD Machines, "ISO 10218 & ISO/TS 15066: Robot Safety Guide"
What we used
ISO/TS 15066 Annex A table of maximum permissible force and pressure by body region (skull/forehead 130 N, face 65 N, chest 140 N, upper arm/elbow 150 N, hand/fingers 140 N, thigh/knee 220 N); quasi-static thresholds typically 40–65% of transient values; PFL cobots typically operate 250–1,000 mm/s; PFL does not make the end effector safe
- humanoid.guide, "Collision Detection in Humanoid Joints"
What we used
ISO/TS 15066 body model of 29 areas in 12 regions, from arXiv 2311.13814; quasi-static limits 65 N (face) to 220 N (thighs and knees); hands and fingers 140 N; permissible relative speed into the chest for a 20 kg effective robot mass 0.24 m/s; hands and fingers 0.67 m/s at 20 kg, 0.75 m/s at 2 kg
- InMotion Global, "ISO/TS 15066 Body Region Force & Pressure Limits Table"
What we used
Spring constants from Table A.3: skull and forehead 150 N/mm, face 75, neck 50, chest 25, abdomen 10, hands and fingers 75, lower legs 60; head and neck are critical zones where the standard does not permit contact
- Aegis AI, "Cobot Safety Just Got Rewritten: What ISO 10218:2025 Means"
What we used
Part 2 grew from 72 pages to 223; validation of force and pressure now mandatory for PFL installations; Class I robot: manipulator mass ≤10 kg, maximum force 50 N, top speed 250 mm/s; Part 1 covers the robot, Part 2 the installed system
- FedLaws, "Collaborative Robot Safety Standards: ISO 10218, OSHA, and CE Marking"
What we used
Every industrial robot must provide emergency stop (category 0 or 1), protective stop (category 0, 1, or 2), and normal stop (category 2); the 2025 revision emphasizes the robot application as a whole, not just the robot system
- Universal Robots, "Technical Specifications UR5e"
What we used
And UR5e User Manual, six rotating joints, 5 kg payload, 850 mm reach, 20.7 kg arm weight, maximum joint speed 180°/s, tool speed about 1 m/s, pose repeatability ±0.03 mm per ISO 9283, force/torque sensor accuracy 4 N, about 200 W under a typical program and 570 W maximum, footprint Ø151 mm
- RoboSathi, "Levels of Automation"
What we used
Sheridan and Verplank's ten levels (1978, MIT Man-Machine Systems Laboratory, for the US Office of Naval Research); level 5 executes if the human approves; the boundary between 4 and 5 is "who approves," between 6 and 7 whether the human sees the action before or after; Bainbridge's Ironies of Automation (1983); a human in the loop is a control only if the human can really say no
- ChMLTech, "AI Agent Autonomy Levels and Human-in-the-Loop Design"
What we used
Parasuraman, Sheridan and Wickens (2000) four functions (information acquisition, analysis, decision selection, action implementation); EU AI Act article 14 requires effective human oversight on every high-risk AI; autonomy set per action by reversibility, impact, detectability, and novelty
- International Federation of Robotics, via NextMSC and GMI (2026)
What we used
66,000 new cobot installations in 2025, 11% of total industrial robot installations; cobot installations up 7% year-on-year; total industrial robot installations 603,000 in 2025, up 11%; 64,542 cobot installations in 2024. Market-size estimates for 2026 range from about $1.27 billion (GMI) to about $7.56 billion (NextMSC), the spread is about what is counted (arms only vs arms plus software and services), so the article uses the installation counts as the hard anchor and states the range