Cobot buying guide 2026: payload, reach, repeatability and the new ISO 10218 safety rules

How to read a collaborative robot arm datasheet without being misled: what payload, reach, repeatability, speed and IP figures really mean, what changed when ISO 10218 was revised in 2025, and how to match an arm to palletising, machine tending, welding or lab work. The analysis draws on 18 cobot records in the RankedRobot database, from a 0.25 kg desktop arm to a 30 kg palletiser.

Last checked 9 October 2026. Spec-based analysis from maker-published figures; we have not tested these robots hands-on.

The short version

A cobot spec sheet is a set of best-case numbers, each measured under different conditions. Payload is quoted at an ideal tool position. Speed is the arm's limit, not the speed it can run beside a person. Repeatability is a statistical test result, not a guarantee of accuracy. The safety picture has also shifted. Since February 2025, the international standard no longer treats any arm as "collaborative" by itself. Only the finished application can be collaborative, and someone has to prove it with a risk assessment.

So buy in this order: define the task and the tool, check payload at the real tool offset, check reach to the furthest point, then look at environment (IP), software and price. Repeatability rarely decides the choice for industrial work. Every arm in our data rated at 5 kg or more publishes 0.1 mm or better.

The 18 arms in our data

We split the records into two groups. Industrial arms carry 5 kg or more and are sold mainly through dealers and integrators. Light and research arms are smaller and often sold online.

Industrial cobots (5 kg payload and above)

*ABB lists the GoFa 10 and GoFa 12 at 12 kg and 14 kg "wrist down" on its product page; its datasheet gives 10 kg and 12 kg. The three Universal Robots g-Series arms were announced on 14 September 2026 and are listed as announced, not shipping.

Light, desktop and research arms

Use the cobot category to filter all records, or compare any two side by side.

Payload versus reach: read the curve, not the headline

The headline payload is the mass the arm can carry with the load close to the flange. Move the load's centre of gravity further out and the allowed mass falls, because the wrist joints must resist a longer lever. Makers publish this as a payload diagram in the manual. If the diagram is not in the brochure, ask for it before you order.

Three habits in the current data show why this matters:

• Conditional payloads. Universal Robots' UR20 datasheet gives 20 kg, then 25 kg subject to boundary conditions set out in the manual. Its collective datasheet does the same for the UR30 (30 kg, or 35 kg under conditions). ABB's higher GoFa figures apply only with the wrist pointing down.
• Names that do not match the rating. On the g-Series page, the UR10g-1750 carries 8 kg as standard and 10 kg extended. The UR17g-1300 carries 15 kg standard and 17.5 kg extended. The number in the model name tracks the extended figure. Plan on the standard one.
• Reach measured to different points. ABB quotes GoFa reach both to the wrist centre and to the flange (950 mm and 1,050 mm for the GoFa 5). Many makers quote one figure without saying which. A 100 mm gap can decide whether you reach the far corner of a pallet.

Your payload budget must include the gripper or tool, any tool changer, cables and the part itself. Grippers and vacuum heads are rarely light, so weigh or look up yours. Size the arm so that this total sits inside the payload diagram at the real tool offset, with margin.

Repeatability and ISO 9283

Repeatability is how closely the arm returns to the same taught point. Accuracy is how closely it reaches a point it was told to go to by coordinates. Datasheets quote repeatability because it is the smaller and more flattering number.

The test method is set out in ISO 9283:1998, still the current edition. ISO last confirmed it in 2021. Training material from metrology maker Hexagon summarises the method. The arm visits five points in a test cube, 30 cycles each. Repeatability is the mean deviation from the centre of the reached positions, plus three standard deviations. The tool load is the full rated load.

What this means for a buyer:

• Look for the standard. Universal Robots states ISO 9283 on its datasheets. ABB's GoFa datasheet quotes 0.02 mm without naming a method. A figure with no named test is hard to compare.
• Regional sheets differ. FANUC's US CRX page gives the CRX-10iA as 0.04 mm. Its European page says up to ±0.05 mm. Use the more conservative figure.
• Repeatability is not path accuracy. For welding seams, dispensing or anything taught offline in simulation, path and absolute accuracy matter more. ABB sells an optional Ultra Accuracy feature for GoFa and quotes path accuracy down to 0.03 mm. Most makers in our data publish no accuracy figure at all.

For most handling tasks, 0.1 mm is ample. Grippers, fixtures and part variation usually contribute more error than the arm does.

Speed: joint limits, TCP speed and collaborative speed

Two speed figures appear on datasheets: maximum joint speeds in degrees per second, and maximum tool centre point (TCP) speed in metres per second. Both are limits the arm can reach with no person nearby.

Be careful with older figures. Universal Robots' e-Series manuals give the UR5e and UR10e tool speed as approximately 1 m/s. UR's 2023 collective datasheet gives 4 m/s as the maximum for both. Our records hold the manual figure. Neither is wrong. One is a typical value and the other a ceiling, and the sheet does not always say which.

ABB's GoFa datasheet adds a footnote buyers should read twice: "Safe collaborative speed will be lower." In a power-and-force-limited application, the allowed speed depends on what part of the body could be struck and how. The result of the risk assessment sets it, not the datasheet. If your business case assumes the arm runs at 5 m/s beside people, it is probably wrong. High speed pays off when the arm runs fenced, or uses speed and separation monitoring to run fast while nobody is near and slow down as someone approaches.

Joint speed also limits cycle time. UR lists 120°/s at the base and shoulder of the UR20 and UR30 and 210°/s at the wrists. ABB lists 120 to 125°/s at the GoFa base. A base-joint swing between conveyor and pallet often decides the cycle, so check the base figure, not only TCP speed.

Force and torque sensing

Every arm sold for collaborative use has to detect contact, but makers do it differently:

• Torque sensors in every joint. ABB says GoFa has integrated torque sensors in all six joints. Doosan says the M1013 has six torque sensors, one per axis. Franka says the Research 3 has seven, one per joint, read at 1 kHz.
• A force/torque sensor at the flange. Universal Robots publishes a full flange sensor spec. On the UR20 it is a ±200 N range, ±5.5 N precision and ±10 N accuracy for force, and a ±20 Nm range for torque. The UR5e's range is 50 N with 4.0 N accuracy.
• Unspecified internal sensing. FANUC describes internal force sensing as standard on every CRX. Techman says built-in sensors stop the arm on contact. Epson describes the AX6 as power- and force-limiting without naming the sensor.

For safety, the method matters less than the certified result. For process work such as sanding, polishing, insertion or screwdriving, the published range and accuracy matter a lot. A sensor accurate to ±10 N cannot control a 5 N contact force. If your task depends on controlled force, ask for the accuracy figure, or budget for a separate wrist sensor.

Environment, weight and mounting

IP rating. The IP code (IEC 60529) has two digits: the first covers dust and solid objects, the second covers water. Higher is better on each. In our data, IP54 is the floor for most industrial arms. Universal Robots describes the IP65 g-Series as suited to washdown environments. FANUC says the IP67 CRX is protected against dust and oil leakage. That matters inside a CNC machine where coolant mist is constant. Note that the controller and teach pendant are often rated lower than the arm. UR lists its UR20 control box at IP44 and pendant at IP54.

Cleanrooms. Epson rates the AX6 for an ISO 14644-1 Class 5 cleanroom. UR publishes an ISO 14644-1 class for each model, with operating conditions attached.

Weight and footprint. Arm weight decides whether you can mount on a mobile cart, a table or a machine. The median weight of the 14 industrial records is 39.6 kg. The UR20 and UR30 weigh 64 kg and 63.5 kg on a 245 mm base and need a rigid pedestal. The Epson AX6 is 18 kg in carbon fibre, and Epson lists tabletop, ceiling and wall mounting. UR and ABB permit any orientation.

Power. Published draw ranges from 36 W (Kinova Gen3) to 750 W maximum (UR20, UR30). The UR10e manual gives both a 615 W average and roughly 350 W with a typical program. Check which figure a sheet means.

Safety in 2026: what changed in ISO 10218

The 2025 revision

ISO published ISO 10218-1:2025 (robots, for manufacturers) and ISO 10218-2:2025 (applications and cells, for integrators) in February 2025. They replace the 2011 editions, which ISO has withdrawn. According to A3's FAQ, the US body that administers US membership of the ISO committee:

• The terms collaborative robot and collaborative operation are gone. The standard speaks of a collaborative application.
• The collaborative requirements formerly in ISO/TS 15066 are now built into the series.
• Safety-rated monitored stop is renamed monitored standstill.
• Functional safety requirements are spelled out rather than implied.

A3 also says the revised Part 1 sorts robots into two classes by mass, maximum force and maximum speed. Class 2 covers most industrial robots and carries extra requirements. A3's news release adds cybersecurity requirements and guidance on end-effectors. In the US these standards are adopted as ANSI/A3 R15.06-2025. A3 notes they are voluntary unless a regulation references them.

ISO/TS 15066: not withdrawn yet

There is confusion here. ISO's own catalogue page for ISO/TS 15066:2016 still lists it as current. In June 2025 ISO changed its stage to to be revised, and named ISO/AWI 15066-1 (Physical contact with robots — Part 1: Biomechanical thresholds and data) as its replacement. That project is at an early stage. In practice, the 2025 ISO 10218 series covers collaborative applications, and the biomechanical limits for power and force limiting are being moved to a dedicated standard.

What the datasheets say

No record in our database names the 2025 edition. Techman cites ISO 10218-1:2011 and ISO/TS 15066:2016. Franka cites the 2011 edition and says certification is still in progress. Universal Robots lists EN ISO 10218-1 with no year. Its Gen 7 announcement says g-Series arms are TÜV certified to ISO 10218-1 and UL 1740, and the safety system is PL d, Category 3 under ISO 13849-1. Ask any supplier which edition its certificate covers.

Risk assessment is not optional

No arm is safe to use beside people out of the box. The integrator must do a risk assessment of the whole application: arm, tool, part, fixtures, and where people stand and move. The general method is ISO 12100, currently being revised. ISO 10218-2:2025 applies it to robot cells. A sharp gripper or a heavy part can rule out power and force limiting even on a fully certified arm. In the EU, the new Machinery Regulation (EU) 2023/1230 applies from 20 January 2027 and replaces the Machinery Directive. Integrators placing cells on the EU market should plan their conformity assessment against it.

Programming and ecosystems

Every arm here offers a graphical interface for hand-guided teaching. The differences show up in what lies beneath.

• Universal Robots: PolyScope on e-Series and UR-Series arms, PolyScope X with open APIs on g-Series, plus URScript and official ROS and ROS 2 drivers. The UR15 and UR30 list Modbus TCP, EtherNet/IP, PROFINET and PROFIsafe.
• FANUC: tablet teach pendant with drag-and-drop icons, ROBOGUIDE simulation and an open-source ROS 2 driver. The main draw is the same R-30iB Plus controller family as FANUC's industrial robots.
• ABB: Wizard block programming on the pendant and RobotStudio for offline simulation.
• Doosan: DART platform and Doosan Robot Language. Techman: TMflow node-based flows with built-in vision. Epson: AX Portal no-code software, a Python environment and a 3D simulator.
• Research arms: Franka offers 1 kHz low-level control via its FCI interface. Kinova offers the Kortex API in C++ and Python with ROS and ROS 2.

Pick the ecosystem your integrator and maintenance staff already know. Tooling marketplaces, spare parts and local support usually outweigh small spec differences.

Choosing by application

Palletising. Payload is the gripper plus the heaviest case, and reach must cover the far corner at the lowest and highest layer. That favours the long-reach arms: UR20 (20 kg, 1,750 mm), UR10g-1750 (8 kg standard, 1,750 mm) and GoFa 10 (1,520 mm to the wrist). For heavier cases at shorter reach, the UR30 carries 30 kg at 1,300 mm. FANUC offers a palletising mode on the CRX. If the stack is taller than the arm's vertical envelope, a lift column is one option.

Machine tending. Coolant and chips make IP rating the first filter. The CRX-10iA, GoFa 10 and 12 are IP67, and the UR g-Series, UR15, UR20 and UR30 are IP65. Reach through the machine door and a small base footprint matter more than top speed.

Welding. Payload must cover the torch and its cable package. Hand-guided teaching speeds up short-run work. ABB and Doosan both list welding among their applications. Path accuracy matters more than repeatability here.

Labs, cleanrooms and research. The Epson AX6 has ISO Class 5 cleanroom rating. The 7-axis Franka Research 3 and the Kinova Gen3 offer low-level control for research.

Education and prototyping. The myCobot 280 and Lite 6 publish online prices but no IP rating and 0.5 mm repeatability. They are good for teaching, but not for production.

Prices: what is actually published

ABB, Doosan, Epson, FANUC, Franka, Kinova and Techman sell on quotation only. Note the delivery terms: the xArm 6 price is ex-works, so the buyer pays import duty, while the Lite 6 price includes it. Grippers, safety devices, integration and the risk assessment all add to the arm price. We have not found a primary source that puts a reliable figure on that share, so get a full cell quote.

What the numbers say (Analysis)

• Repeatability is no longer a differentiator for industrial arms. All 18 records publish a figure. Across all 18 the median is 0.065 mm (range 0.02 to 0.5 mm). Among the 14 arms rated at 5 kg or more, the median is 0.05 mm and the worst is 0.1 mm.
• Prices are scarce. Only 6 of 18 records carry a price: three maker list prices (Elephant, UFACTORY ×2) and three distributor-reported Universal Robots prices. Seven are quote-only, and five newer UR arms have no price on record.
• IP65 is becoming normal at the top. 15 of 18 records publish an arm IP rating. 8 reach IP65 or better on at least one variant, including all six UR arms launched from 2022 onwards. The three lowest-priced arms publish no rating.
• Payload-to-weight varies sixfold. Among the 17 records with a confirmed weight, payload per kilogram of arm ranges from 0.075 (Lite 6) to 0.47 (UR30), with a median of 0.31. The UR18g-950 (0.46) and UR30 trade reach for payload. The UR10g-1750 (0.18) trades payload for reach.
• 1,300 mm is the default reach. 6 of 18 records list exactly 1,300 mm (Doosan M1013, Techman TM12, UR10e, UR15, UR17g-1300, UR30). Only 3 exceed 1,500 mm (GoFa 10, UR10g-1750, UR20).
• Open software is standard. 15 of 18 records list ROS or ROS 2 support (Doosan's from a media report). The exceptions are ABB GoFa, Epson AX6 and Techman TM12.

For a ranked view of advanced robots across categories, see Robot500. Our best cobots list and cobot buying guide hub are updated as records change.

Frequently asked questions

Can a cobot work without any guarding?

Sometimes. It depends on the risk assessment of the whole application, not the arm. ISO 10218:2025 no longer describes the arm itself as collaborative. A sharp tool, a heavy part or a pinch point against a fixture can mean you need guarding or scanners even with a certified arm.

Has ISO/TS 15066 been withdrawn?

Not as of October 2026. ISO lists it as current but due for revision, with ISO/AWI 15066-1 on biomechanical thresholds named as its replacement. A3 says its collaborative requirements are now part of the 2025 ISO 10218 series.

What repeatability do I need?

For pick-and-place, palletising and machine tending, 0.1 mm is usually more than enough. For precision assembly, look for 0.05 mm or better, check that the figure is stated to ISO 9283, and ask about absolute and path accuracy.

Why does the payload on the product page differ from the datasheet?

Higher figures often apply under conditions: wrist pointing down, a load close to the flange, or reduced speed. Plan with the standard figure and check the payload diagram at your actual tool offset.

Is the maximum TCP speed usable beside people?

Usually not. ABB states that safe collaborative speed will be lower than maximum. The allowed speed comes from the risk assessment and the power-and-force limits for the body regions at risk.

How much does a cobot cost?

Published prices in our data run from USD 599 (myCobot 280 M5) to a reported USD 62,990 (UR20). Most industrial makers quote only. Add grippers, safety devices and integration to get a real budget.

Do I need ROS?

Not for production. Graphical pendants handle most industrial tasks. ROS or ROS 2 support matters for research, custom vision or fleet integration, and 15 of our 18 records list it.

Methodology

Specifications come from the RankedRobot database. Each figure is taken from a maker product page, datasheet or manual, and each robot page lists its source and whether a figure is maker-stated, reported or unconfirmed. Distributor prices are marked as reported. Standards status was checked on the ISO catalogue, A3 publications and EU-OSHA on 9 October 2026. Statistics use the 18 cobot records live on that date; for multi-variant records (ABB GoFa) we used the largest listed payload when computing medians. We have not tested these robots hands-on; this is spec-based analysis. See our methodology and use the robot finder to filter by payload, reach and IP rating.

Spotted an error or a newer figure? Email corrections@rankedrobot.com and we'll check it and log the correction.