Micropsi Industries teaches robot skills for variable manufacturing tasks

Micropsi Industries develops MIRAI, software that learns robot skills from demonstrations and uses camera feedback to handle variation during execution. The buying decision is whether a difficult operation can become a reusable skill within a commissioned manufacturing cell.

01 / ProductMIRAI adds learned guidance to an industrial robot

Micropsi Industries describes MIRAI as a combination of robot-control software, a learning service and a training application. People demonstrate the motion needed for a task; the system turns those examples into a skill that guides the robot from visual feedback. The current product page presents MIRAI 2 and support for KUKA, FANUC and Universal Robots. Specific robot and controller compatibility still needs confirmation for the proposed cell.

The company is selling a way to manage variation at a particular step. A part may arrive slightly displaced, a cable may bend differently, or a reflective component may be difficult to locate using a fixed visual rule. Learning the relationship between camera input and motion can reduce dependence on perfectly repeated presentation. That is an engineering hypothesis to test against the actual part population, rather than a reason to remove all fixtures.

MIRAI 2 adds semi-automatic recording, sharing of skills and detection of abnormal conditions, according to the product description. It also removes the need for a force-torque sensor in most skills. That qualification matters: a reader should not carry forward older installation descriptions as universal requirements, or interpret “most” as a promise that every contact-sensitive operation needs no additional sensing.

02 / AudienceStart with the variable operation, not the whole factory

The application catalogue covers jobs such as insertion, end-of-line testing, picking, screwdriving and leak detection. These are bounded physical operations with a recognizable beginning and end. An appealing candidate is one where the surrounding process already works, but a changing part pose or flexible item forces manual intervention. The useful question is how much of that variation can be represented through demonstrations.

An automotive engineering team might consider a skill for guiding a tool onto a fastener or handling a flexible component. Micropsi’s automotive page explicitly discusses such variability. A proposed trial should include difficult reflections, awkward orientations and part tolerances rather than only the cleanest sample. The trial result would determine whether the workpiece presentation or tooling also needs to change.

The company’s about page positions manufacturing as the established application area while discussing broader possibilities. That makes a production engineering team a clearer starting audience than a buyer seeking a general warehouse robot. MIRAI supplies intelligence for a robot application; the buyer still needs the arm, end effector, machine interfaces and a responsible integrator.

03 / WorkflowA proposed skill begins with demonstration and ends with a controlled handoff

Consider a proposed insertion cell in which an operator currently compensates for small changes in component position. First, the team defines the accepted part variants and the state in which the robot may start. It then records demonstrations covering the intended range of approach angles, lighting and displacement. The point is to teach the correction that the variable step needs, while keeping the rest of the process understandable.

Next, the training workflow computes a skill and the team executes it on representative workpieces. Micropsi describes an iterative process of recording, computing, executing and refining on its product page. A useful evaluation separates ordinary successes from cases that require additional examples, altered tooling or a different task definition. It should also record which variants are deliberately outside the accepted operating envelope.

The company’s older vision-system explanation describes handing control between the native robot program and MIRAI for the variable part of an operation. That remains a useful conceptual distinction, while current hardware requirements should come from the current product specification. The surrounding program must decide when to call the skill, what counts as completion and what happens after a stop.

For example, an abnormal-condition stop could leave a part partially inserted. Automatically retrying from that state may be a different physical task from starting with a free part. The proposed commissioning plan should therefore define removal, inspection and restart behavior explicitly. Sharing a successful skill between cells can reduce repeated training work, but each installation still needs its own verification of camera placement, tooling and workpiece presentation.

04 / PricingLicensing changes with who deploys the skill

The buying page distinguishes several commercial routes rather than posting a single subscription price. A company with one application can work with Micropsi and an integration partner. A corporate agreement covers unrestricted deployment counts and application types within that company’s internal manufacturing. The distinction is the right to use the software across sites, not proof that every site will need identical engineering effort.

Micropsi also describes source-code licensing for automation equipment makers and a separate system-partner route. The former supports white-label products and derivative work; the latter describes perpetual, worldwide, royalty-free rights for delivered machines after purchase. These are materially different from an internal manufacturing license. A machine builder should make its distribution and support obligations explicit before treating software rights as reusable across customer projects.

The page says licenses include a software support agreement covering updates, cloud services and user training. It also lists Micropsi-operated cloud service, a customer AWS tenant and local deployment options. No numerical tariff is published on the page we read. The quotation should distinguish the permitted deployment model, continuing support and the hardware or integration work needed to make the first cell productive.

05 / DistinctionsThe distinction is learned motion inside an existing automation stack

MIRAI is relevant when detecting an object is only part of the problem. The robot must also adjust its motion as the visual situation changes. Cognex provides a useful adjacent comparison for teams whose main requirement is inspection or locating an item. The decision is whether the job ends with a measurement or requires a learned relationship between perception and physical movement.

FANUC is a complementary comparison because the robot platform and its surrounding controls remain part of the installation. A manufacturer should evaluate MIRAI against the difficult step it may simplify, while preserving a clear account of what the robot controller, PLC, machine and operator each do. The benefit is easier to assess when the proposed skill has a narrow operational boundary.

Skill sharing is particularly relevant for a manufacturer with repeated processes in multiple factories. A common demonstration and refinement process could create a more reusable engineering asset than a separate hand-tuned correction at each station. That is our interpretation of the product model, not a measured savings claim. Differences between cells may still make adaptation and local acceptance work substantial.

06 / QuestionsResolve the limits of a learned skill before duplicating it

The most consequential uncertainty is generalization: which changes resemble the demonstrations closely enough for reliable execution? A buyer can make that question concrete by holding out realistic part variants during training and testing them later. The evaluation should report the accepted range and the recovery behavior, rather than one pooled success percentage that obscures a troublesome subgroup.

Deployment location also changes the operating relationship. If training or other services run in a cloud environment, the agreement should identify the data transmitted, the customer’s control over stored demonstrations and what remains usable during a connection interruption. The published alternatives establish that deployment choices exist; they do not specify every security or offline-operation term for an individual contract.

Finally, clarify responsibility for maintaining a deployed skill after a camera, gripper or component changes. The attraction of learned behavior can weaken if nobody owns requalification. A useful handover includes the task boundary, the representative training set, known exclusions and a procedure for deciding when a change calls for new demonstrations.

07 / DecisionChoose the commercial route that matches the deployment owner

Micropsi Industries is worth investigating when a manufacturer can name the variable motion that blocks otherwise practical automation. A focused trial can establish whether demonstration-based guidance handles that variation and whether the remaining cell engineering is economical. Its different licensing routes also make it relevant to equipment builders, provided internal use and onward distribution are treated as separate decisions.