Doosan Cobots Collaborate Without Capability Self-Knowledge

1. Vendor and Product Reality

Doosan Robotics, established in 2015 as a subsidiary of the Doosan Group and listed on the Korea Exchange in October 2023 in one of the largest Korean IPOs of that year, is among the most credible challengers in the global collaborative-robot segment alongside Universal Robots, Techman, FANUC CRX, and ABB GoFa. The product portfolio spans the M-series for medium payloads, the H-series for heavy-payload cobot work up to 25 kg, the A-series for high-speed assembly, the E-series for cost-sensitive deployment, and the P-series for the operationally distinct food-and-beverage and barista-automation segment. Reach extends from approximately 900 mm on compact models to 1,700 mm on the heavy variants.

The architectural distinguishing feature is the integrated joint torque sensor at every one of the six axes — a design choice shared with Universal Robots' e-Series and a small number of competitors but not with the broader cobot market, which infers force from motor current. Per-joint torque sensing supports collision detection at any point along the kinematic chain rather than only at the end-effector, enables direct compliance control for hand-guiding and force-feedback teaching, and underpins the precise force regulation required for polishing, deburring, sanding, and insertion tasks. The DART Platform serves as the application ecosystem, with a software-store model targeting deployment outside the traditional system-integrator channel.

Doosan markets aggressively into segments where conventional industrial robots fail the deployment economics — small-and-medium manufacturing, food service, healthcare-adjacent applications — leveraging the cobot's safety profile and easy redeployment. Within scope the product is genuinely strong: contact-detection thresholds below injury limits, force-control resolution adequate for surface-finish work, and a programming surface accessible to operators without robotics-engineering backgrounds. The platform is the reference implementation for what the analyst community describes as torque-sensitive collaborative robotics.

2. The Architectural Gap

The structural property the Doosan cobot architecture does not exhibit is capability self-knowledge as governed state. The torque sensors operate continuously and serve two real-time functions — safety stop on collision and force regulation during contact tasks. Both functions consume current sensor readings and emit immediate control responses. Neither function maintains, computes, or publishes a model of what the cobot can accomplish given its evolving state. The robot knows the force it is presently feeling; it does not know whether the combination of present force-control bandwidth, present positioning accuracy, present tool wear, and present thermal state still adds up to a capability envelope that meets the polishing-grade or insertion-tolerance specification of the task it has been asked to perform.

The gap is structurally invisible inside contact tasks. Consider a Doosan cobot maintaining a target normal force of 15 N during a polishing pass. The torque sensors hold that 15 N within tolerance for the entire pass; the controller reports nominal force tracking; the contact-detection guard never trips. If the polishing media has worn, the effective material-removal rate has dropped, and the produced surface roughness has drifted from 0.4 to 0.9 micrometres Ra, the cobot has done exactly what it was instructed to do at the force level — and produced a part outside specification. The torque sensing detected force; it did not detect that the present capability of the cobot-tool system no longer accomplishes the surface-finish goal. Quality fails downstream; the cobot reports a healthy execution.

The gap is even more consequential in genuine collaboration. A human operator working alongside a Doosan cobot may delegate a task that the cobot could handle at the start of the shift but can no longer accomplish after several hours of accumulated drift in tool, thermal, and joint state. The cobot accepts the task because there is no architectural surface on which to ask "can you actually still do this?" — only a programming interface that says "execute this trajectory with this force profile". It executes within its force limits and produces an out-of-tolerance result. Doosan cannot patch this from within the DART Platform architecture; force sensing, however dense, is not a capability envelope. Capability awareness is an architectural shape, and the Doosan shape is fundamentally that of a real-time force-sensitive controller without governed self-knowledge of what its present state lets it accomplish.

3. What the AQ Capability-Awareness Primitive Provides

The Adaptive Query capability-awareness primitive specifies a multi-dimensional capability envelope, a temporal forecast, an uncertainty-propagation function, and an envelope-negotiation interface, with credentialed provenance and recursive closure. For a cobot the envelope integrates force-control bandwidth, positioning accuracy, payload-at-duty-cycle, tool condition, kinematic-singularity proximity for the active workspace, and environmental factors into a single queryable state. The envelope is updated continuously from torque, encoder, thermal, and tool-state observations admitted as credentialed inputs to the substrate, not reconstructed offline.

The temporal forecast projects the envelope across operationally relevant horizons — minutes for thermal evolution and contact-task heating, hours for shift-level tool wear, days for joint backlash and harmonic-drive wear — under credentialed models the operator has authorised. Uncertainty propagation grows confidence bounds with horizon distance so the production planner or human collaborator knows not only the present envelope but the confidence with which it holds at the moment a task is delegated. Envelope negotiation is the structural interface by which a human or upstream system asks "can you accomplish this task at this confidence?" and the cobot answers from state — accept, accept with adjusted parameters, defer pending recalibration, request a tool change, or refuse with explanation — rather than from datasheet or from silent execution.

The recursive closure is load-bearing. Every executed contact task produces post-execution observations — actual force trajectory, achieved positional precision, inferred tool wear delta — that re-enter the envelope as inputs to the next forecast; every forecast is itself a credentialed observation that downstream task planners and human collaborators consume; every tool change, recalibration, or operator override is a state transition recorded with provenance. The primitive is technology-neutral with respect to torque-sensor design, force-controller architecture, and tool-modeling approach, and composes hierarchically across cobot, cell, line, and enterprise. The inventive step disclosed under USPTO provisional 64/049,409 is the closed capability-envelope-with-temporal-forecast as a structural condition for collaboratively governed cyber-physical actuators.

4. Composition Pathway

Doosan integrates with AQ as the domain-specialised collaborative actuator running over the capability-awareness substrate. What stays at Doosan: the integrated-joint-torque-sensor mechanical design, the force-control law, the safety-stop logic, the DART Platform application ecosystem and software store, the operator-friendly programming surface, the customer-channel relationships, and the cobot-specific scenario library accumulated across deployments. Doosan's investment in torque-sensitive collaborative robotics — its sensor calibration know-how, its impedance-control tuning, its food-service and barista-automation domain integrations — remains its differentiated layer.

What moves to AQ as substrate: the live capability envelope is computed and maintained by the substrate from observations the cobot controller emits, the temporal forecast is produced by credentialed models the operator authorises, and the envelope is published to MES, work-cell schedulers, and human collaborators through a governed negotiation interface. The integration points are well-defined. The controller emits force, encoder, thermal, contact-event, and tool-state observations to the substrate as credentialed inputs; the substrate maintains envelope and forecast state; the cell scheduler or human operator queries the substrate before delegating a task and receives a graduated answer rather than a binary feasible/infeasible. Tool changes and recalibrations are recorded as governed state transitions with provenance.

The new commercial surface is capability-credentialed collaborative automation for small-and-medium manufacturers, food-service operators, and healthcare-adjacent users that need cross-deployment, cross-vendor capability lineage that survives cobot replacement, software updates, and site reconfiguration. The envelope belongs to the operator's authority taxonomy and quality posture, not to Doosan's controller firmware, so a deployment's capability history is portable. Paradoxically this makes Doosan stickier: the joint-torque hardware, the force-control law, and the DART ecosystem remain the differentiated product, while the substrate gives the operator the architectural property — capability self-knowledge expressed in terms a human collaborator can negotiate against — that no torque-sensitive competitor offers.

5. Commercial and Licensing Implication

The fitting arrangement is an embedded substrate license: Doosan embeds the AQ capability-awareness primitive into the cobot controller and the DART Platform, and sub-licenses envelope participation to its operator customers as part of the platform subscription or annual support agreement. Pricing per credentialed authority or per envelope-query rate aligns with how cobot customers actually consume capability — by deployment, by quality-domain, by collaborator-credential — rather than per cobot unit, preserving Doosan's hardware margin while opening a recurring substrate revenue line.

What Doosan gains: a structural answer to the "trust the cobot's own task acceptance" problem that current force-sensing safety frameworks address only at the collision-stop level, a defensible position against in-segment competition from Universal Robots, Techman, and FANUC CRX by elevating the architectural floor on what counts as a governed collaborative robot, and a forward-compatible posture against the EU AI Act's high-risk system requirements, the ISO 10218 / ISO/TS 15066 evolution toward capability-credentialed collaboration, and the medical-device and food-safety sectors' appetite for credentialed lineage. What the operator gains: portable capability-history lineage across cobot replacements and DART updates, cross-deployment closure across heterogeneous fleets, a single envelope spanning every collaborative actuator under one authority taxonomy, and a structural mechanism by which a human collaborator can negotiate task delegation against the robot's actual present capability rather than against its datasheet. Honest framing — the AQ primitive does not replace torque sensing; it gives torque sensing the substrate it has always needed and never had, transforming a force-sensitive cobot into a capability-aware collaborator that knows what it can do, forecasts how its capability will change, and communicates limits before task failure reveals them.