Safety and Compliance for Hobby Robot Arms

Who is asking decides which rules apply

The most common way people get this question wrong is to start from the arm. They look at a printed thing with six bus servos sitting on a desk, decide it is obviously harmless, and stop. Then they look at the same arm in a company, decide it is obviously a machine, and panic. Both readings are wrong, because EU product law does not classify objects by how dangerous they look. It classifies events: placing a product on the market, and putting it into service.

So the first question is not what the arm is. It is who is holding it, who paid for it, and who else is in the room. An SO-100 you printed and drive yourself at your kitchen table, an identical SO-100 on a bench in a company where a colleague walks past, and an SO-100 you assemble and sell for 400 EUR are three different legal situations with the same bill of materials. Nothing about the servos changed. What changed is that in the second case there is an employer and a worker, and in the third there is a supply.

This is also why the answers you find online are so unhelpful. Somebody asks whether their hobby arm needs CE marking and gets two confident replies: no, because it is a toy, and yes, because it is machinery. Both are answering a question about the object. The useful answer starts from what the person intends to do with it, and for most readers it is short. The table below is what I would tell a colleague before they read a single article number, and for a lot of people the first row is the entire answer.

What the Machinery Regulation actually says

Regulation (EU) 2023/1230 was adopted on 14 June 2023 and applies on a mandatory basis from 20 January 2027, replacing Directive 2006/42/EC. The European Commission is explicit that machinery placed on the EU market before 20 January 2027 must comply with the old Directive. That is worth internalising, because a lot of writing on this implies the Regulation retroactively changes what is already on a bench. It does not. If you are working through a build right now, the SO-100 hub and the complete SO-100 setup guide are the practical companions to this page; what follows is only the part about who is allowed to do what.

Placing on the market is not the same as putting into service

This distinction is the whole game for anyone building their own arm, and it is where most confident internet answers fall apart. Making available on the market is defined as supply in the course of a commercial activity, which a private build clearly is not. Putting into service carries no such qualifier: it is simply the first use for its intended purpose in the Union. The European Commission's guidance on the Machinery Directive has long been read to mean that a person who builds a machine for their own use has not placed it on the market but has put it into service, and therefore carries the manufacturer's obligations.

Read literally, that would sweep in every hobbyist. In practice it does not, because market surveillance under EU product law is aimed at economic operators, and because the enforcement hook for own-use machinery in every member state I have looked at runs through workplace legislation rather than product legislation. The honest formulation is this: the risk of a compliance problem is essentially zero while the arm stays private, and becomes real the moment it is at a place of work. Do not build a business plan on the first half of that sentence.

• Private, own use, own home: no realistic obligation. Nobody is coming.
• Own build, own company, your own employees use it: you are the manufacturer for the purposes of putting into service, and you are also the employer under Directive 2009/104/EC. The second duty is the one with inspectors attached.
• Bought as a kit, assembled per instructions: assembling per instructions is not a substantial modification. You are a user.
• Bought as a kit, then you bolted on a bigger gripper and a 12 V supply: that is exactly the shape of a substantial modification under Article 3(16), and it moves manufacturer duties to you.
• Sold to someone else, even at cost, even to a friend: supply in the course of a commercial activity is read broadly, and free of charge is explicitly included.

The machine-learning clause that everyone misreads

The headline change in the Machinery Regulation for anyone running a policy on hardware is in Annex I Part A. Two of the six high-risk categories that require a notified body concern machine learning: safety components with fully or partially self-evolving behaviour using machine learning approaches ensuring safety functions, and machinery that embeds such components. This gets quoted constantly as proof that AI-driven machines now need third-party assessment. That reading skips the most important five words in the sentence.

• Ensuring safety functions. The clause is about software that performs a safety function: stopping the machine, blocking access, limiting speed. A vision-language-action model that decides where to move the gripper is a control function. It is what the machine is for, not what keeps it safe.
• Self-evolving. A checkpoint you trained once and froze does not evolve in the field. It produces different outputs on different inputs, which is not the same thing. Continual on-robot learning is a different conversation.
• Both conditions of the AI Act must be met. Article 6(1) of Regulation (EU) 2024/1689 classifies an AI system as high-risk only if it is a safety component of a product covered by Annex I legislation and that product must undergo third-party conformity assessment. Fail either limb and you are out.
• The hobbyist exemptions are broad. AI Act Article 2(8) excludes research, testing and development activity prior to placing on the market or putting into service, though not testing in real world conditions. Article 2(10) excludes natural persons using AI systems in a purely personal non-professional activity. Article 2(6) excludes systems developed solely for scientific research and development.

So the practical answer for someone fine-tuning ACT or SmolVLA on their own recordings is that the AI Act does not reach them, and the machine-learning clauses of the Machinery Regulation do not either. The checkpoint is the productive part of the machine. It decides where the gripper goes, well or badly, and when it decides badly the remedy is a better dataset, not a notified body.

Where it does change is if you build a safety layer out of a model. A vision model that decides whether a person is close enough to stop the arm is a safety component in the literal sense the Regulation means: it is ensuring a safety function, and if it keeps learning in the field it is self-evolving as well. That is the case that pulls you into Annex I Part A and third-party assessment, and it is also the case where the AI Act's two-limb test in Article 6(1) is finally satisfied on both limbs. The engineering advice is blunt: do not build that. Use a light curtain, a two-hand control, an interlocked enclosure or a switch. They are cheap, boring, assessable, and nobody has to argue about a detector's confidence threshold after an incident.

The standards, and why none of them quite fit

Once you accept that some duty applies, the next instinct is to reach for a standard. This is where a small arm gets awkward, because the robotics standards were written for two worlds and a desk arm belongs to neither. The table is the map.

The practical consequence of that table is worth stating plainly, because it surprises people who expect a standard to exist for every product. There is no harmonised standard written for a sub-2 N.m desk manipulator, and there probably never will be, because the category is too small and too heterogeneous to justify one. That is not a loophole and it is not a gap you can hide in. Where a duty applies, the absence of a type C standard means you carry the risk assessment yourself under ISO 12100 rather than presuming conformity by following a checklist. In exchange, nobody can tell you your end effector fails clause 5.10.3 of a standard that was written for a 200 kg industrial arm behind a fence.

The physics: what 1.6 newton metres can actually do

Compliance arguments are more useful when you know the numbers. The SO-ARM100 build repository states that the 7.4 V STS3215 used in an SO-100 follower has a stall torque of 16.5 kg.cm at 6 V, and that a 12 V variant of the same part number exists at 30 kg.cm. Converting the 7.4 V figure: 16.5 kg.cm is about 1.6 N.m. At 20 cm from the joint axis that is roughly 8 N, about the weight of a 800 g object. At 30 cm it is closer to 5 N. Stall torque is a brief peak, not something the servo sustains, and the SO-100 follower uses six identical 1/345 gear servos, so no joint is dramatically stronger than another. The SO-101 uses the same 7.4 V servos with mixed gear ratios, and the SO-100 against SO-101 comparison sets the two side by side. A Koch v1.1 is a different electrical animal, with 5 V and 12 V rails, and its numbers do not carry over.

That is genuinely low. It is not enough to break a finger by blunt impact, and it is well below the biomechanical thresholds that collaborative-robot work concerns itself with. Stall torque is also the wrong number for impact: what injures is force concentrated over a small area, and this arm funnels its modest torque into a gripper jaw a few millimetres wide. Pressure, not force, is why the gripper is the hazard on this machine and the links are not.

Low force is not the same as no hazard, and the hazards that recur on this hardware are not the ones people brace for. The swinging link is not the problem. The recurring incidents are a finger in the gripper because somebody reached in to reseat an object mid-episode, a cable dragged off the bench by an unclamped base, and an arm going limp while holding something over a laptop. All three are workspace problems. All three are solved with clamps and habits, not with law, and none of them appear anywhere in a conformity assessment.

The three software limits you actually have

LeRobot gives you three mechanisms that behave like safety measures, and it is worth knowing exactly what they do and what their defaults are, because two of the three are off unless you turn them on. These come from calibration and from the follower robot config, read from the LeRobot main branch on 2026-08-24. If you have not calibrated an arm yet, the SO-100 getting started walkthrough and the teleoperation documentation cover the mechanics; this section is only about which of those knobs behave like limits.

1. 1
   Cap how far a joint can move in one step

   max_relative_target clamps the difference between the commanded position and the measured present position, per motor. It defaults to None, which means no clamp at all. Set it to a float for all joints or a dict per joint. The cost is real: enabling it forces an extra sync_read of the follower every step, and the source comment says to expect slower fps.

   bash

   # clamp every joint to 5 units of travel per control step
   lerobot-teleoperate \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431551 \
     --robot.id=my_follower \
     --robot.max_relative_target=5 \
     --teleop.type=so100_leader \
     --teleop.port=/dev/tty.usbmodem58760431552 \
     --teleop.id=my_leader

2. 2
   Let calibration set the joint travel limits

   Calibration records range_min and range_max for each motor and writes them to the servo. This is the only limit that is on by default, and it is the reason a badly calibrated arm can drive a link into a hard stop. If a joint stops short of where you expect, that is calibration, not the policy.

   bash

   lerobot-calibrate \
     --robot.type=so100_follower \
     --robot.port=/dev/tty.usbmodem58760431551 \
     --robot.id=my_follower
   
   # move to mid-range, press ENTER, then sweep every joint
   # except wrist_roll through its full travel

3. 3
   Know which torque limits are written, and why

   At connect time LeRobot writes PID gains to every motor and torque limits to exactly one: the gripper. The comments in the source say these exist to avoid burnout. They are motor protection, not human protection. The five arm joints run at the servo's own defaults.

   python

   # lerobot/robots/so_follower/so_follower.py, configure()
   self.bus.write("P_Coefficient", motor, 16)
   self.bus.write("I_Coefficient", motor, 0)
   self.bus.write("D_Coefficient", motor, 32)
   
   if motor == "gripper":
       self.bus.write("Max_Torque_Limit", motor, 500)   # 50% of max torque
       self.bus.write("Protection_Current", motor, 250) # 50% of max current
       self.bus.write("Overload_Torque", motor, 25)     # 25% when overloaded

Beyond that, the Feetech register table LeRobot uses exposes limits you can write yourself if you want a harder ceiling than the defaults. These are servo registers, not LeRobot features, so nothing in the stack manages them for you and nothing resets them when you unplug.

# Registers exposed in lerobot/motors/feetech/tables.py (STS3215)
Min_Position_Limit        addr 9    joint travel floor
Max_Position_Limit        addr 11   joint travel ceiling
Max_Temperature_Limit     addr 13   thermal cutoff
Max_Voltage_Limit         addr 14   over-voltage guard
Min_Voltage_Limit         addr 15   under-voltage guard
Max_Torque_Limit          addr 16   ceiling written to gripper only (500)
Protection_Current        addr 28   written to gripper only (250)
Overload_Torque           addr 36   written to gripper only (25)
Torque_Limit              addr 48   runtime torque scaling
Present_Temperature       addr 63   read-only, watch this one
Present_Current           addr 69   read-only

When it stops being a toy

There is no line in any regulation that says a robot arm becomes serious at a given mass or torque. But there are thresholds in practice, and every one of them is about context rather than the machine. Here is the list I would actually use. It is the same instinct as the one behind collecting high-quality training data: the thing that changes the answer is almost never the hardware.

• Somebody who did not build it operates it. Your tacit knowledge of where the arm swings is not transferable. This is the single biggest step change.
• It runs unattended. A policy executing rollouts while you make coffee is a different machine from one you are watching.
• It leaves the bench. A mobile base changes the reach envelope from fixed to unbounded. A LeKiwi is not a desk arm.
• The payload gets heavier or sharper. Force at the gripper is force at the gripper. What it is holding decides the consequence.
• It is at a place of work. Employer duties do not scale with machine size, and an inspector does not care that it cost 130 EUR.
• You sell one. The transaction is the trigger, not the price.
• You bolt on a bigger gripper, a 12 V supply, or a second arm. Look again at Article 3(16) on substantial modification.

Liability does not wait for CE marking

Everything above is product law and workplace law: rules about whether you may put a thing on the market or into use. There is a third body of law that people forget entirely, and it is the one most likely to matter to a hobbyist, because it attaches to harm rather than to paperwork. Directive (EU) 2024/2853 on liability for defective products has to be transposed into national law by 9 December 2026 and replaces the 1985 directive that has governed this since before any of us were writing robot code.

Two changes in it are directly relevant to anyone running a policy on hardware. First, the definition of product now explicitly covers software, standalone or in combination with another product. A trained checkpoint is software. Second, free and open-source software is carved out of scope when it is not supplied in the course of a business activity, and that carve-out disappears if there is a chargeable service, or a service supplied in exchange for personal data.

A risk assessment you can finish in an afternoon

If you decided a duty applies, this is the work. It is not a certification exercise and it does not need a template from a consultancy. ISO 12100 is the type A standard that describes the method, and the method is four moves: determine the limits of the machinery, identify the hazards, estimate and evaluate the risk, then reduce it in a fixed order. The fixed order is the part people get wrong. Inherently safe design beats guarding, guarding beats a warning label, and a warning label on its own is almost never an adequate answer to a hazard you could have designed out with a clamp.

1. 1
   Fix the limits of use in writing

   Reach envelope in centimetres, maximum payload, operating hours, who may operate. Take a photo of the bench and mark the envelope on it. If you cannot draw the envelope, you do not know it.

2. 2
   Walk the hazards physically

   Power the arm, drive it slowly through its full travel with the leader arm, and watch what it comes near. Do this once with the bench as it actually is, not as you tidied it for the photo. Add anything you find to the hazard list.

3. 3
   Apply the three-step method in order

   Inherently safe design first, then guarding and protective devices, then information for use. Clamping the base is step one. A warning sticker is step three, and a step-three control alone is almost never enough.

4. 4
   Test the stop you actually have

   Confirm what happens on Ctrl-C, on unplugging USB, and on cutting power. On a stock LeRobot config, all three end in a limp arm. Decide whether that is acceptable at every height the arm reaches, and if it is not, do not let it reach that high with a payload.

5. 5
   Write the two-page record

   Hazards, controls, residual risk, who checked and when. That is the whole document. If an employee uses the arm, this is not optional and it must exist before first use.

   text

   SO-100 bench arm, hazard record, 2026-08-24
   
   Limits    reach 35 cm radius, payload under 200 g,
             operator: named individuals only
   Hazards   gripper pinch; trap against clamp; limp drop
             on disconnect; hot servo on stall
   Controls  base clamped x2; envelope cleared 40 cm;
             nothing breakable under raised arm;
             switched power strip within arm's reach of
             the operator; TPU compliant fingers fitted
   Residual  limp drop cannot be eliminated: no brake
             exists. Height limited by procedure.
   Review    on any gripper or payload change

Where AY-Robots does not help

This platform records LeRobot datasets, fine-tunes five policies on rented GPUs and serves the result back to an arm. None of that is a compliance product. It cannot CE mark anything, it does not know what is on your bench, and there is no emergency stop in the stack because there is no hardware in the stack that belongs to it. A policy that fails safely is a property of your wiring and your workspace, not of the checkpoint.

One more honest limit. Nothing on this page is legal advice, and the reason it can be this specific is that the underlying question is usually simpler than people fear. If your arm is private, you are fine. If it is at work, do the hazard assessment, and do it before the first person who is not you presses go. If you sell arms, get a real compliance review, because from 20 January 2027 the file you are building is against a Regulation and not a Directive. Everything between those three cases is detail. And if none of this appeals, start without owning an arm and let the machinery duty stay with somebody who already has a risk assessment on file.