Agility Robotics' Digit Walks Without Knowing What It Can Carry

1. Vendor and Product Reality

Agility Robotics, spun out of Oregon State University and now operating from Salem, Oregon with a dedicated humanoid manufacturing facility (the RoboFab) capable of producing thousands of units annually, is one of a small handful of companies shipping commercially deployed bipedal humanoid robots into customer-paid logistics workflows. Digit, now in its fourth-generation industrial form factor, has been deployed by Amazon, GXO Logistics, Spanx, and other warehouse operators in pilots and early production for repetitive bin and tote movement tasks — recirculation of empty totes from outbound to inbound, decanting from gaylord containers, and tote-to-conveyor transfer. The published payload is roughly sixteen kilograms; battery life runs in the multi-hour range with hot-swap capability; locomotion handles flat industrial flooring and modest thresholds and ramps.

The engineering is real and unusual. Digit's leg architecture combines spring-loaded passive dynamics with high-bandwidth electric actuation, and its locomotion stack draws on more than a decade of bipedal control research from the Cassie research platform. Its perception stack handles tote identification, fiducial reading, grasp planning for standard warehouse tote geometries, and obstacle avoidance in shared-floor environments. The Agility Arc cloud platform manages fleet operations, task assignment from warehouse-management systems, and remote monitoring. The customer pitch is structurally honest — Digit is a humanoid because warehouses are built for humans, and bipedal locomotion plus humanoid arm geometry lets it occupy infrastructure that AMRs and shelf-shuttles cannot.

Within its current scope, Digit is the most operationally credible bipedal humanoid in commercial logistics. The competitive set — Figure, Apptronik, 1X, Tesla Optimus, Boston Dynamics' Atlas successor — is moving quickly, but Agility is further into customer-paid deployment and has the advantage of a logistics-specific product focus rather than a general-purpose humanoid pitch.

2. The Architectural Gap

The structural property Digit's architecture does not exhibit is a persistent, integrated capability envelope that the robot maintains as a first-class state variable, exposes to its task planner and to the warehouse-management system, and forecasts forward in time. Today's stack is layered the conventional way: a low-level locomotion controller maintains balance under current conditions; a manipulation controller plans grasps; a battery-management subsystem reports state of charge; a thermal-management subsystem reports motor and electronics temperatures; a fleet-management layer assigns tasks. Each subsystem is competent in isolation, but there is no architectural object that integrates them into a published, queryable answer to the question "what can this robot reliably accomplish in the next thirty minutes given its current and projected state?"

The gap matters because warehouse tasks live in the cross-product of many capability dimensions. A heavy tote on a long route with a moderate ramp at end-of-shift battery level with elevated motor temperatures from the previous hour's duty cycle is a task whose feasibility is not a function of any one subsystem in isolation; it is a function of their interaction. The current architecture handles this implicitly: the controller attempts the task and either succeeds with reduced margins or fails when combined demands exceed capability. Failure mid-task in a humanoid form factor — a stumble under load, a dropped tote, a stop on a ramp with insufficient remaining battery to recover — has a much worse cost profile than the equivalent failure in an AGV.

Agility cannot patch this from within its current control stack because the capability envelope is not a controller function — it is a representation function that has to live above the controllers and integrate their state into a deterministic, forecastable, negotiable object. Adding more health telemetry, more conservative defaults, or more aggressive learning policies does not produce capability awareness in the architectural sense; those are tactical mitigations of an absent representation.

3. What the AQ Capability-Awareness Primitive Provides

The Adaptive Query capability-awareness primitive specifies that a conforming cyber-physical system maintain a persistent capability envelope as a first-class state object, with five structural properties. First, the envelope is multi-dimensional and explicit: locomotion capacity, manipulation capacity, perception confidence, energy reserve, thermal headroom, and any additional domain dimensions are represented as named state variables with defined ranges and units. Second, the envelope is integrated: cross-dimension dependencies are expressed as published functions, so locomotion margin is reduced as payload increases, manipulation reach is reduced as base support narrows under load, and reliable carrying capacity is reduced as surface friction degrades.

Third, the envelope is forward-projected: temporal forecasting produces an expected envelope trajectory over a defined horizon under current task assumptions, so the system can answer not only "what can I do now" but "what will I be able to do in twenty minutes if I continue this duty cycle." Fourth, the envelope is negotiable: when a task assignment falls outside the current or projected envelope, the system emits a structured gap report — payload exceeds capacity by 1.8 kilograms given current battery and the ramp on the assigned route — that an upstream planner can resolve by reassignment, route change, or scheduled recovery. Fifth, the envelope is auditable: envelope evaluations and the events that updated them are recorded as lineage so post-incident reconstruction can show what the robot believed it could do at any past moment and why.

The primitive is technology-neutral — any underlying controller stack, any battery chemistry, any sensor suite — and composes hierarchically: per-robot envelopes roll up to fleet envelopes under the same structural rules. The inventive step is the unified, forecastable, negotiable representation, not any one capability sensor.

4. Composition Pathway

Agility Robotics integrates with AQ as a domain-specialized humanoid platform running over a capability-awareness substrate. What stays at Agility: the locomotion controller, the manipulation stack, the perception system, the mechanical design, the RoboFab manufacturing capability, the Agility Arc fleet platform, and the customer relationship. Agility's investment in bipedal locomotion is its differentiated layer and remains so.

What moves to AQ as substrate: the capability envelope object, its cross-dimension integration logic, its temporal forecasting, its negotiation protocol with upstream planners, and its lineage recording. The integration points are well-defined. Subsystem telemetry — battery state, motor torque headroom, thermal state, controller margin, perception confidence — is published into the envelope under the primitive's update rules. The Agility Arc task assignment path is mediated by the negotiation protocol: tasks arrive as proposed assignments, the envelope evaluates feasibility under current and projected state, and the response is either accepted, accepted with conditions (reduced speed, alternate route), or refused with a structured gap. The warehouse-management system sees a robot that publishes what it can do rather than a robot that reports what it has done.

The new commercial surface is capability-governed humanoid logistics. Operators gain predictable behavior at the edges of the envelope, where today's deployments are conservatively throttled to avoid the failure modes that an integrated envelope makes plannable. Insurers and safety regulators gain audit-grade lineage of what the robot believed it could do at the moment of any incident, which directly supports the certification regimes — ISO 10218, ANSI/RIA R15.08 mobile manipulators, emerging humanoid-specific frameworks — that are converging on self-assessment requirements.

5. Commercial and Licensing Implication

The fitting arrangement is an embedded substrate license: Agility embeds the AQ capability-awareness primitive into Digit's onboard software and into the Agility Arc fleet layer, and sub-licenses substrate participation to operators as part of the robot-as-a-service subscription. Pricing is per-deployed-unit-hour or per-envelope-evaluation rather than per-task, which aligns with the operational reality that the envelope is evaluated continuously, not per work order.

What Agility gains: a structural answer to the operator question "how do I plan around a humanoid that I cannot fully predict," a defensible position against well-funded competitors whose stacks are equally controller-centric and equally lacking an integrated envelope, and a forward-compatible posture against emerging humanoid-safety certification regimes that will require demonstrable self-assessment. What the operator gains: predictable utilization at higher envelope occupancy, plannable battery-swap and recovery scheduling, audit-grade incident reconstruction, and a fleet-level envelope view that supports cross-vendor operations as more humanoid form factors enter the floor. Honest framing — the AQ primitive does not replace bipedal locomotion engineering; it gives that engineering the integrated self-knowledge it has always needed and never had.