Best Cobots and Industrial Arms: Production Line or Research Bench

There is one question that settles most of this category before a single specification matters: is the arm going into production, or onto a research bench. A production arm is optimised so the people who work beside it can re-task it, and so the vendor publishes the figures you will be held to in a specification document. A research arm is optimised for low-level access, torque control at the joint, and reproducing published work. The two lists of virtues barely overlap, and buying one machine for the other machine's job goes badly in ways that show up months later.

So the ranking criterion here is fit to that split, weighted by how honestly each vendor supports the buyer with published data and a maintained software path. Underneath that sits the second thing that decides projects: rated payload covers everything past the flange, and it derates with reach and orientation. Get that wrong and the arm you bought cannot do the job you bought it for, whichever side of the split you are on.

Companion reading: how to choose a cobot, collaborative robots (cobots) ultimate guide, and the full reviews behind this roundup: Universal Robots, Franka Research 3, KUKA LBR iiwa.

Table of contents

1. Key takeaways
2. The ranking at a glance
3. Universal Robots
4. Franka Research 3
5. KUKA LBR iiwa
6. How to choose between them
7. Verdict
8. Frequently asked questions
9. Changelog

The ranking at a glance

Repeatability is deliberately absent from that table. UR and Franka publish theirs, the FR3 at ±0.1 mm. KUKA does not publish one for the iiwa at all, and the figures that circulate are third-party.

Universal Robots

Best at: production cells that have to change. The hardware is good and no longer unique, and competitors match or beat the specs and often undercut the price. What UR still owns is that a maintenance technician can be taught to program one. The teach pendant, the waypoint model and the largest third-party accessory catalogue in the category mean the cell gets re-tasked by the people standing next to it rather than by a contractor on a callout. Over five years that moves the cost of automation more than the purchase price does. The range spans 3 kg at 500 mm on the UR3e up to 30 kg on the UR30 and 1750 mm on the UR20, so most bench and light material-handling tasks land somewhere inside it.

The one disqualifying caveat: the abstraction that makes it approachable is also its ceiling. If the application needs 1 kHz torque control, custom impedance behaviour or a learned policy, UR's interface is the constraint and a research-grade arm is the right tool. The secondary trap is model selection. Read the lineup as one ladder and you can end up with an arm that lifts the part and cannot reach the far side of the fixture. Fix reach against the real cell first, including the approach posture, then check payload against the full tool stack.

Full review: Universal Robots Review: The Lineup Is Two Ladders, Not One.

Franka Research 3

Best at: manipulation research, and teaching that should involve force rather than only waypoints. Torque sensors sit at every one of its seven joints and are exposed through the Franka Control Interface at 1 kHz for torque commands and sensor readback, so an impedance controller written on it is a real impedance controller. The seventh axis gives redundancy: the arm can hold the tool pose fixed while reconfiguring its elbow around an obstacle, which a six-axis arm cannot do. ROS and ROS 2 integration, MATLAB and Simulink support, and the fact that the papers you want to reproduce were written on this arm compound into a network effect worth months. Repeatability is quoted at ±0.1 mm, which is industrial-grade precision on a research-grade machine.

The one disqualifying caveat: 3 kg over an 855 mm reach, and that budget covers the gripper, its mounting and the cabling before the workpiece. With a Franka Hand or a third-party gripper fitted, the part allowance is small. Read it as a scope statement: 3 kg is enough for essentially every published manipulation benchmark and not enough for production material handling. The corporate history is the second thing to carry into a decade-long platform decision. Franka Emika filed for insolvency in August 2023 after failing to secure follow-up financing, citing irreconcilable shareholder differences, and Agile Robots AG acquired it in November 2023 in a deal reported above €30 million, retaining the roughly 100-person team. It trades as Franka Robotics today. Roughly 700 patents were sold a few weeks before the filing, the administrator worked to reverse that sale and shareholders obtained preliminary injunctions against the purchase agreement, with the final position not clearly reported, and the Munich I public prosecutor's office opened a subsidy-fraud investigation. The product is staffed, sold and supported, which beats liquidation. Get the spare-parts commitment and support horizon in the quote.

Full review: Franka Research 3 Review: The Lab Default, and Its Corporate Asterisk.

KUKA LBR iiwa

Best at: contact-rich research that needs seven-axis redundancy with genuine joint torque sensing at a higher payload than the FR3 offers. KUKA's product page describes joint torque sensors in all seven axes and an axis-specific torque accuracy of ±2% of the maximum torque, with payload of 7 to 14 kg, reach of 800 to 820 mm, IP 54, and ceiling, floor and wall mounting (floor only for cleanroom). The hardware is still excellent, and used and refurbished units circulate on the secondary market, which is the cheapest way for a lab to find out whether the arm suits the work.

The one disqualifying caveat: the software path is a dead end and the data is missing. The LBR iisy runs iiQKA.OS, KUKA launched iiQKA.OS2 in 2025, and in April 2026 trade coverage reported a new KUKA controller shown at Global Industrie framed around coming EU cybersecurity mandates. The iiwa is on none of those roadmaps. It stays on Sunrise.OS in a Sunrise Cabinet that integrator documentation and university setup guides consistently describe as Windows 7 Embedded plus a virtualised VxWorks, and Windows Embedded Standard 7 lost extended support in October 2020. That description could not be confirmed against a KUKA document, and whether units shipping in 2026 use a newer Windows build is an open question worth putting to a sales engineer in writing. Zero published CVEs name the LBR iiwa, Sunrise.OS, Sunrise Cabinet or Sunrise.Workbench, which reads as an absence of audit rather than an absence of bugs; the sibling KR C4 and KR C5 stack carries CVE-2022-2242 at CVSS 9.8, unauthenticated read and write of robot configuration over TCP port 49003. The Fast Robot Interface, the reason most research buyers pick this arm, is a separately licensed technology package sitting behind the my.kuka.com login wall with no public pricing, and no authentication for it is documented in KUKA material or in the open-source drivers (unverified rather than confirmed absent). Add the missing figures: no repeatability, weight, speed or price on the public page. RoboDK's entry for the 14 R820 lists 0.1 mm repeatability and 30 kg weight, and those are third-party numbers that could not be traced to a KUKA document, so treat them as unsourced until KUKA supplies one in writing.

Full review: KUKA LBR iiwa Review: The Cobot KUKA Stopped Modernising.

How to choose between them

Answer the split first. If the arm has to earn its keep on a line, and your own staff will modify the program next year, it is UR, and the remaining work is model selection. Treat reach and payload as separate constraints, measure the furthest point the tool must reach in the real cell including the approach posture, pick the smallest arm that covers it, then verify payload against the whole tool stack at your worst-case reach and orientation.

If the arm is an instrument, the question becomes what you need below the position interface. For torque control at 1 kHz, reproducibility against the published literature and a bundled real-time interface, the FR3 is the default and the default is worth something. If you need seven-axis dexterity at 7 to 14 kg and you have Java capability in house through Sunrise.Workbench, the iiwa is the arm that fits the envelope, on the condition that you can run it on an isolated network and that the licence line items are itemised in the quote.

Two disqualifiers to apply before anything else. A cell that must still be compliant and supported in 2028 argues against the iiwa: the EU Machinery Regulation 2023/1230 applies from 20 January 2027, and trade coverage of EN ISO 10218:2025 reports that it addresses cybersecurity expectations for connected manufacturing environments. And any specification you will be held to argues against a machine whose vendor will not publish the number. If your process depends on absolute positioning off a CAD frame rather than relative moves off a taught frame, budget for a calibration campaign and measure the result on your own arm, because repeatability and absolute accuracy are different quantities and the gap is large on serial arms of this class.

Then budget the cell instead of the arm. On UR, accessories, integration and training can roughly double the base arm price, and on arms in the iiwa's class integration routinely rivals the hardware line. A business case built on the arm price alone understates the project by around half.

Verdict

Universal Robots takes the top slot because it is the right answer for the largest share of buyers and it makes the fewest demands on the organisation around it: published figures, a bundled real-time interface, the deepest accessory catalogue in the category, an integrator network that has seen your application before, and programming a technician can learn in days. Its limits are honest and visible on the way in.

The Franka Research 3 ranks second on the same test applied to a different job. For manipulation research it is the best-supported default there is, and the 3 kg ceiling is a scope statement. The corporate turbulence is real and mitigated, and it converts into diligence questions. The LBR iiwa is third on software and disclosure. The hardware is not what holds it back. Seven axes with torque sensing on all of them at up to 14 kg is genuinely capable, and it sits on an operating system with no announced successor, driven by a real-time interface that costs extra and that KUKA suggests you air-gap. That is survivable in a lab at used prices and hard to defend in a 2027 factory audit.

Frequently asked questions

Which of these three should I buy? Decide production versus research first. Universal Robots for production cells your own technicians re-task, the Franka Research 3 for manipulation research needing 1 kHz torque control, the KUKA LBR iiwa for seven-axis contact-rich research at 7 to 14 kg where you can accept the Sunrise.OS situation.

Does the rated payload include the gripper? Yes, and more. The rating covers everything past the flange: gripper, mounting plate, tool changer, the cabling and services the arm carries, and the part. Centre of gravity and inertia count too, especially on fast moves or when the tool sticks far out from the flange.

Why does effective payload drop as the arm extends? Payload derates with geometry. Effective capacity falls as the arm extends and as tool orientation moves away from the favourable case. The clearest evidence is that some UR variants publish a higher rating specifically for wrist-down applications such as palletising.

Can I use a UR arm for research? Up to a point. For 1 kHz torque control, custom impedance behaviour or learned policies, a research-grade arm like the FR3 is the right tool. UR's teach-pendant abstraction becomes the constraint at that level.

Why is the LBR iiwa ranked below the FR3 when it carries more? Software path and disclosure. UR and Franka publish their performance figures and bundle the real-time interface. KUKA's public iiwa page omits repeatability, weight, speed and price, FRI is separately licensed behind a login wall, and the iiwa appears on no iiQKA roadmap while KUKA's other cobot line moves on.

Is the LBR iiwa discontinued? No end-of-life or last-order-date notice was found, and it is still listed as a current product on kuka.com as of August 2026. It is absent from KUKA's iiQKA operating-system roadmap.

How much of the project cost is the arm? On UR, accessories, integration and training can roughly double the base arm price, so the arm is roughly half the project. On arms in the iiwa's class, integration routinely rivals the hardware line, and the Sunrise licence packages sit on top at an undisclosed price. Cost the cell, then check that the arm is the cheap part.