Flexiv builds force-sensitive robots for contact-rich industrial work

Flexiv’s adaptive robots combine force control, visual perception and AI for jobs where contact with the world matters. Inserting a connector or finishing a surface requires more than replaying a path: the robot must accommodate what actually happens at the tool. This review explains the documented platform and a proposed evaluation, without claiming a hands-on test.

01 / ProductAdaptivity centers on force as well as position

Flexiv’s adaptive-robot explanation describes force control and visual perception working together to tolerate positioning error and disturbances. The company’s AI relevance is physical: software must convert observations into useful motion and contact. That does not make every operation a learned policy or remove the need to specify the task.

The Rizon family supplies seven-axis arms, a controller and supporting accessories. The current page lists Rizon 4 and Rizon 10 configurations, including versions with an additional force-torque sensor. Elements runs on the teach pendant for visual task programming. Mounting, tooling, camera integration and communication interfaces belong to the system decision, not merely the arm selection.

Flexiv also offers robot systems: MICO is built on Enlight for dual-arm work, while AICO uses Rizon and is positioned for embodied-intelligence development. These are distinct configurations, not interchangeable product names. The company’s overview also includes Moonlight and GRAV in its portfolio. A buyer should ask which hardware and software combination supports the intended application.

02 / AudienceThe strongest fit involves useful, measurable contact

Assembly teams should investigate Flexiv when a small variation in part pose makes an otherwise straightforward operation unreliable. A connector might need a controlled search before insertion, or a tool may need to maintain contact as a surface changes. In those situations, reducing fixture precision requirements could be valuable, but only if completed-part quality remains acceptable.

An embodied-AI research group has a different requirement: consistent observations and physical control through repeated experiments. MICO and AICO provide a hardware-system conversation, while access to interfaces, supported control modes and data capture needs its own verification. A published robot-control interface does not establish that a particular research model is already integrated or commercially supported.

For a conventional production cell, the Teradyne blueprint offers context on collaborative-arm platforms and application ecosystems. For general robot-policy research, the Physical Intelligence blueprint addresses a different layer of the stack. Flexiv supplies robots and control capabilities; a model developer’s broader learning ambition should not be treated as an out-of-the-box Flexiv application.

A weak fit is a project that cannot describe the workpiece, allowable contact, acceptance condition or recovery behavior. Adaptive control can respond to variation, but it does not decide the business definition of a good assembly. Those definitions must come from the manufacturing or research team.

03 / WorkflowA proposed insertion trial should expose the failure modes

Consider a proposed Rizon trial for inserting a connector into a housing. Start with a small family of real parts and the actual quality requirement. The purpose is to determine whether the system accommodates normal presentation differences while avoiding damaged contacts, incomplete seating and excessive cycle-time variation. Do not judge it solely by successful motion on a carefully prepared demonstration part.

The Primitives Manual describes actions as building blocks for larger PLANs. It lists assembly primitives such as SearchHole, CheckPiH, InsertComp, Mate and FastenScrew, alongside motion and force-control actions. These names provide a concrete starting point for an application discussion; they do not prove that every listed action suits every connector or installed software version.

Ask the integrator to explain the proposed sequence: approach, establish contact, search if necessary, insert, verify and release. Define what each stage observes and how it terminates. A contact threshold may indicate that something was touched, but it is not by itself proof that the desired feature was found. A useful trial makes the transition conditions visible to the team reviewing the result.

Vary tray placement and part presentation within a documented envelope. Include a missing connector, a blocked housing and a part that is already seated. The expected result for an unsuitable input may be a controlled stop and an intelligible exception, rather than a completed insertion. Counting that distinction prevents a superficial success metric from rewarding damage or unnecessary retries.

The manual also requires appropriate robot calibration, tool-weight accuracy and operation within joint and workspace limits. Check these conditions before attributing a failed trial to the learning or force-control method. For example, a tool configuration that changes between runs can make otherwise comparable contact measurements difficult to interpret. The test record should identify which configuration produced each result.

Inspect the finished assemblies using the process’s normal quality method. Report accepted parts per operating window, incomplete insertions, interventions and any damage. Include part loading and recovery effort rather than measuring only the robot’s active movement. This is a proposed evaluation design, not a published Flexiv throughput result.

For a research system, add a separate reproducibility check. Repeat the same task with the same software and hardware configuration, preserve the observations available to the controller, and distinguish training demonstrations from held-out trials. A dual-arm platform can make data collection convenient while still leaving the researcher responsible for the evaluation split and the definition of generalization.

04 / PricingA complete adaptive cell needs a scoped quotation

Flexiv’s contact page provides business inquiry routes and regional sales contacts. The inspected pages do not display a universal arm price, subscription tariff or installed-cell amount. Budgeting should begin with a configuration and application discussion rather than a price inferred from a different distributor package.

The Rizon page describes the Hesper controller, teach pendant, motion bar, end-of-arm tooling and camera options. Establish which items are included and which are optional in the quoted system. The robot’s published capability is not a statement that every accessory arrives in a base package. A workstation also needs the physical arrangement and interfaces that make its task possible.

Separate production integration from research enablement. A manufacturing proposal should identify the delivered task and acceptance work, while a research quote should identify available interfaces, development access and support. Public documentation is useful for technical investigation, but it does not establish a license entitlement, maintenance term or guaranteed support level.

05 / DistinctionsThe documentation makes contact behavior discussable

A useful distinction is that Flexiv exposes an application vocabulary through its primitive documentation. The buyer can ask how a search, contact or insertion is configured instead of accepting only a general promise of intelligence. That vocabulary can make an application easier to inspect, although documentation alone cannot establish how much tuning a specific process needs.

Seven-axis motion can provide flexibility when choosing a posture around a fixture, but the extra axis is not a universal cure for difficult access. The relevant test is whether the complete tool and workpiece can reach the needed pose while preserving a useful force-control posture. Fixture design and cable routing remain part of that question.

The integrated systems also offer an interesting bridge between industrial control and embodied-AI experiments. A research team can investigate perception and learning on hardware designed to execute physical contact. The value depends on the supported interfaces and the quality of the experimental setup, rather than simply the presence of two arms or an AI label.

06 / QuestionsSoftware versions and mechanical constraints remain consequential

The cited primitive manual is explicitly version 3.11 for Rizon 4(s). Match documentation to the delivered robot and software release before adopting a parameter or action. Other model and version links exist, so readers should not assume that a familiar primitive name implies identical behavior across all configurations.

Public pages advertise tolerance to positioning errors, but an application still has a finite capture range and acceptable force envelope. Ask for those boundaries on the representative part family. If the supplier cannot show what happens outside them, a successful demonstration gives only a narrow answer about the demonstrated setup.

The current system page mixes several MICO and AICO configurations with different physical arrangements. Confirm availability, accessories and the supported development environment for the exact system offered. This review establishes public product and inquiry routes; it does not establish inventory, delivery dates, universal regional service or performance on the proposed connector task.

07 / DecisionChoose Flexiv when contact quality is the actual problem

The next useful step is a representative task study with visible completion criteria. Flexiv’s force-control approach deserves attention where contact variation consumes engineering or operator effort, while broader AI research needs a separate agreement about interfaces and experimental control.