Dexterity turns physical AI into robotic handling for logistics

Dexterity develops AI and robotic systems for handling physical goods in enterprise logistics. Its offer connects perception, packing decisions, motion planning and force control with industrial hardware, including its Mech robot. The buyer’s task is to judge an operational handling system: the number of packages moved correctly, the interruptions it requires and how it fits an existing flow of goods.

01 / ProductPhysical reasoning connects to industrial manipulation

The platform page describes an integrated stack of skill agents, hardware abstraction, safety monitoring and telemetry. Its named applications include trailer loading and unloading, palletising, depalletising and singulation. Those applications share a difficult property: objects arrive with variation, and moving one object changes what the robot can safely do next.

Foresight is Dexterity’s world model within that stack. The company describes maintaining a changing representation of the physical scene, considering possible actions and updating the representation after contact. That is relevant to manipulation because an intended placement and an observed placement can differ. A box can settle, a grasp can slip and another surface can become visible.

Mech is the company’s industrial robot, while IRIS is its hardware abstraction layer for supported arms and mobile platforms. These names describe different layers of the offer. A customer should establish which combination is proposed for its site. Software portability does not mean a deployment can use arbitrary existing equipment without validation or integration.

Dexterity’s company history describes its origins in 2017 and expansion into enterprise logistics, including work in Japan. Its current homepage names relationships with FedEx, Sumitomo and Sagawa Express. These establish the operational setting the company targets; they do not establish that every facility or advertised application has the same deployment maturity.

02 / AudienceHigh-volume handling sites have the clearest evaluation problem

A parcel hub, distribution centre or logistics operator can investigate Dexterity when workers repeatedly handle a varied stream of goods. The relevant business problem is usually a specific handling bottleneck: receiving mixed cartons, forming stable loads or moving items into a downstream process. The robot must fit the available space, timing and item mix at that point.

For a plant engineering team, the most useful starting material is a representative description of the work. Include the incoming object distribution, awkward packaging, upstream arrival patterns and downstream acceptance rules. A machine that handles common boxes well may still require an economical route for unusual items. The exception path belongs in the process design from the start.

An AI research team may instead be interested in physical world modelling and online packing. Dexterity has made part of that problem visible through its public challenge documentation. That material can help explain the reasoning problem, but it should not be mistaken for an open commercial robotics API with a self-service deployment route.

A small business looking for an off-the-shelf robot that handles unrelated household or office chores is outside the core buying situation shown by the sources. Dexterity’s value proposition depends on industrial workflows, repeatable volume and an implementation relationship. The sophistication of the AI does not remove that operational context.

03 / WorkflowA proposed trailer-loading evaluation follows the whole load

Consider a proposed evaluation at one loading dock. Sequenced has not operated Dexterity hardware. Begin by defining the supported trailers and package classes, where packages enter the cell and what a completed acceptable load looks like. A loading robot is evaluated within a flow: an upstream conveyor, the robot, the forming load and the departure schedule all affect the result.

Collect a baseline over representative shifts. Measure handled packages, loading time, damaged goods, rework and interruptions, while retaining the volume and package distribution for each shift. Then run the proposed automated process against comparable conditions. Avoid comparing a robot’s easiest sequence with a human team’s busiest, most irregular period.

The Foresight explanation shows why placement decisions involve more than fitting rectangular volumes together. Stability, reachability and the consequences of later actions matter. In this proposed trial, inspect the finished load and its handling requirements as well as package throughput. Higher density is useful only if the load remains acceptable for the actual transport operation.

Track interventions by cause. Separate an unsuitable package from poor sensing, an unavailable conveyor, a failed grasp or a site procedure that leaves the cell waiting. This helps identify whether the next improvement needs a model change, hardware adjustment or a different operating routine. A single blended uptime number can hide those very different problems.

Close the trial with the operators responsible for ordinary shifts. Can they identify the state of the cell, recover approved exceptions and return to production without ambiguous workarounds? The proposed result should be a supported operating envelope and a measured improvement against the existing dock process, not a claim that physical handling has been solved universally.

04 / PricingDeployment economics are negotiated around a particular operation

The current contact route invites enterprise deployment discussions. The reviewed platform and contact pages do not publish a universal price per robot, package or hour. A buyer therefore needs a scoped offer covering the proposed application. It would be misleading to turn a research challenge prize or a vendor investment announcement into product pricing.

The 2026 Foresight API challenge explicitly states that it has concluded. Its retained documentation describes a packing competition and historical academic eligibility. The continued presence of API examples does not prove that a new team can obtain access today, and competition access was never evidence of rights to control production hardware.

Ask the commercial team to separate equipment, integration, support and any recurring software or service charges in the offer. The public sources do not establish which contracting structure will apply. For financial comparison, relate the full cost to accepted goods handled and usable operating hours, including the labour and equipment that remain around the automated cell.

Capacity and density can create value in different ways. Completing a trailer earlier matters only if it changes a real constraint, while fitting more goods can affect transport utilisation. Build the business case from the site’s own constraints. Vendor statements about overall production activity are evidence of experience, not a forecast of savings at another facility.

05 / DistinctionsThe application is specialised even when the intelligence is reusable

Dexterity’s useful distinction is the connection between software that interprets physical change and machinery designed for industrial work. An action has to be geometrically possible, mechanically achievable and acceptable to the process. This makes the integration of planning and execution central to the offer; a correct textual explanation of how to load a box would accomplish none of those things by itself.

The Figure blueprint provides context for humanoid robots aimed at work in environments built for people. Dexterity frames Mech around logistics-specific manipulation. That difference helps a buyer ask whether human-like form is essential to its task or whether a purpose-built arrangement can better fit the work area.

The Boston Dynamics blueprint offers another comparison around mobile robotics and commercial applications. Compare the exact job, supported goods and surrounding process rather than treating every company that moves objects as an interchangeable supplier. A trailer-loading decision and a facility-inspection decision can share AI techniques while having completely different acceptance criteria.

The challenge documentation also makes a helpful distinction explicit: parts of its evaluation excluded robot reachability, visibility and observation noise. Those omissions make a research problem tractable. They also explain why a high packing-game result cannot be read as demonstrated industrial robot performance.

06 / QuestionsAsk which claims apply to the proposed installation

Dexterity reports substantial autonomous production activity on its homepage. Such statements are vendor claims and have not been independently audited by Sequenced. Before using them in a procurement model, ask how an action is counted, which applications contribute and what period the figure covers. An action count is not the same unit as accepted packages, operating hours or completed trailers.

Hardware compatibility needs the same precision. Ask for the supported robot, gripper, sensing arrangement and software release in the proposed system. A general hardware abstraction story is useful architecture, but a specific customer needs a qualified configuration and someone accountable for keeping it working when components change.

Availability should be established application by application. A platform page can list a family of uses without telling a buyer whether each is standard, newly offered or subject to further engineering. Request the current deployment scope and representative evidence for the chosen use. That prevents an existing loading capability from being stretched into an unsupported claim about every form of material handling.

Finally, clarify what happens when the physical state is uncertain. The practical questions are how the system pauses, how an operator learns what went wrong and how the scene is restored before work continues. Those details determine whether advanced autonomy becomes a manageable production process.

07 / DecisionBuy an accepted handling outcome

Dexterity is a relevant physical AI company for readers evaluating automation at a logistics bottleneck. Its combination of world modelling, coordinated skills and industrial machinery addresses the difficult space between sensing an object and completing a useful movement. The evaluation should remain anchored to a particular flow of goods and a realistic package mix.

The next step for an operator is a scoped deployment conversation supported by its own operating data. The next step for a researcher is to distinguish the public packing material from production system access. Both can learn from the same technology without confusing a research interface, an advertised application and a commercially accepted installation.