Robot Cell Commissioning is Needed to Produce High Quality Parts
Physical installation of robot cells is only the first step in bringing a robotic cell into production. After the robot, fixtures, tools, sensors, and safety boundaries are physically installed, the system must determine how the actual cell differs from its planned configuration. It must identify the equipment and objects present, estimate their locations and relationships, calibrate the robot, tools, and sensors, update collision models, and verify that each capability can be executed reliably. It must also configure and validate auxiliary behaviors required for autonomous operation, such as replacing a worn sanding disc or interacting with tool and consumable stations.
Cell commissioning is a critical engineering step needed to produce good quality parts from AI-powered robotic cells at customer sites. Today, commissioning requires substantial site-specific engineering work, including constructing an accurate representation of the installed cell; calibrating robots, tools, and sensors; adapting models to local conditions; onboarding production parts; configuring customer-specific operational skills; and testing and validating the complete system before production begins. Historically, much of this work remains manual and dependent on experienced robotics engineers, commissioning has become a major constraint on deployment speed and scalability. GrayMatter Robotics is addressing this bottleneck through AI agents powered by physics-informed world models and foundation models. These agents autonomously build, validate, and maintain the operational knowledge required by each cell, empowering human productivity across every deployment.
Autonomous Commissioning Requires Sophisticated Reasoning
Installation differences can shift fixture and sensor poses; observations may be incomplete or noisy; calibration may drift; previously modeled objects may move; and site-specific structures may create unanticipated visibility, reachability, or collision constraints. An autonomous commissioning system must therefore determine what it knows, identify what is missing, or inconsistent, and select sensing or interaction actions that resolve those uncertainties before the cell begins operation.
The same challenge applies when onboarding production parts. For each new part, the system must determine the part type, its geometry, pose, material and surface characteristics, and relationship to surrounding fixtures. It must then determine which surfaces require processing, how the part should be scanned, which robot poses provide sufficient coverage, and whether the available observations are accurate enough to support planning and execution. The process recipe (Refer to our Whitepaper on AI Agent for Process Recipes) and operational workflow must then be configured and validated against the customer’s production requirements.
Collectively, these activities build and maintain an operational representation of the cell: what is present, where it is located, how confidently it is known, which interactions are possible, and whether the system is ready to operate safely and reliably. This representation provides the foundation for downstream robotic capabilities. Motion planning depends on accurate geometry and collision constraints; scan planning depends on sensor poses, visibility, and coverage; task execution depends on the identities, states, and affordances of tools, fixtures, and parts; and process execution depends on the part, production requirements, and operating context.
GrayMatter Robotics is developing the Cell Commissioning Agent (CCA) to autonomously perform commissioning tasks. Rather than requiring engineers to manually author and maintain this knowledge for every deployment and production part, CCA enables the cell to perform reasoning required to make and keep itself operational. This reduces the engineering effort required per deployment and allows the same deployment and applications teams to support more cells and greater part variety.
Capabilities Needed to Realize Cell Commissioning Agent
CCA uses a closed-loop process that combines scene understanding, uncertainty evaluation, predictive modeling, physical interaction, and validation. CCA first constructs a representation of the cell using the Foundation Model for Scene Understanding. It then evaluates that representation against the requirements of the capability being commissioned to identify information that is missing, stale, inconsistent, or insufficiently reliable. An uncertain fixture pose, for example, may prevent collision-free motion planning; incomplete part geometry may prevent process-path generation; and insufficient knowledge of a consumable station may prevent the robot from reliably replacing a worn sanding disc. CCA resolves these uncertainties by selecting sensing, scanning, calibration, or manipulation actions expected to provide the missing information or validate the required capability. The Sensor World Model supports this reasoning by predicting the observations and information quality expected from candidate sensor actions under the geometry, material, lighting, dust, contact, and other conditions present in manufacturing environments. For vision-based sensing, these predictions help CCA select scan poses that achieve the required coverage and measurement quality. For force and contact sensing, they help the agent select interactions and learn contact-based behaviors such as removing and replacing a sanding disc.
After executing an action, CCA compares the observed result with the expected outcome, validates whether the relevant operational requirement has been satisfied, and updates the cell’s representation. If unresolved uncertainty remains, the loop continues until the required readiness criteria are met. The same mechanism supports initial cell activation, part onboarding, calibration recovery, equipment changes, and continuous commissioning throughout the cell’s operating life. Together, these capabilities transform commissioning from a manually executed checklist into a goal-directed process of uncertainty reduction and validation. CCA does not merely automate individual setup tasks; it determines which commissioning actions are required, executes them through the physical cell, and verifies that the system is ready for production.
What the Cell Representation Contains
The manufacturing cell’s scene representation that CCA builds is produced by the Foundation Model for Scene Understanding, which detects and segments different components within the cell and represents the relationships between objects, regions, sensors, fixtures, tools, and parts as a scene graph. This vocabulary is manufacturing-specific rather than open-vocabulary, since a factory differs substantially from general household or open-world scenes. Inputs across modalities are encoded into this representation, since a manufacturing cell relies not only on vision but also on contact-heavy processes. The representation is hierarchical rather than a flat one, since different hierarchies serve different tasks over the same underlying cell state instead of forcing every downstream task to consume the same view of the scene. A structural hierarchy, for example, can support collision-model updates, while an interaction hierarchy can support affordance reasoning over which tool, fixture, sensor, or region is relevant to a given operation. Fidelity is also non-uniform across the graph: entities such as the part being processed may carry high-fidelity geometry or dense point-cloud detail, while static structures serving only as collision geometry may be represented at coarser resolution, so a downstream consumer can use only the level of detail its task requires.
Commissioning is not a one-time event in high-mix manufacturing
In traditional manufacturing, commissioning is often treated as a discrete phase that ends once a robotic cell has been installed, calibrated, and validated for a fixed production configuration. High-mix manufacturing breaks this assumption. New part variants, fixture changes, tooling updates, consumable stations, and site-specific deviations continually introduce information the cell has not yet represented or validated. CCA therefore treats initial commissioning and later adaptation as instances of the same underlying process. The difference is not in the mechanism, but in the scope of uncertainty. During initial activation, much of the cell is unknown, so CCA broadly establishes fixtures, work zones, sensor relationships, collision geometry, and other operational knowledge. Once the cell is running, the same process is applied selectively whenever a region of the representation requires update. When a new fixture type is introduced, or a new consumable station is added, CCA updates only the affected portion of the cell’s representation. For example, introducing a new sandpaper holder requires the system to identify the holder, estimate its location, determine the pickup and disposal geometry, learn or validate the required interaction, and incorporate that capability into the cell’s operational state. The same evaluate-interact-validate-update loop can be used to recover from calibration drift, or adapt to a changed fixture. By making this process autonomous and continuously available, CCA allows the cell to remain operational as its environment and production requirements evolve.
Commissioning Workflows in Practice
We evaluated CCA on two deployment workflows that previously required direct engineering effort: onboarding a production part family and learning a contact-rich auxiliary skill.
The first workflow involved a customer production part with five SKU variants. Each variant had to be scanned with sufficient coverage and geometric fidelity to support downstream process planning. Before CCA, an engineer traveled to the customer site and manually taught scan poses for this part family. Because overhead lighting affected scan quality, the poses also had to be repeatedly tested and adjusted across parts, days, and times of day. Configuring and validating the complete part family required approximately 2 engineering days on site.
With CCA, the Foundation Model for Scene Understanding identified the part, fixture, support surfaces, and process-relevant regions. The Sensor World Model predicted the scan quality expected from candidate poses under the site’s lighting and environmental conditions. CCA then evaluated the sampled poses and selected a reachable, collision-free pose that satisfied the required scan-quality criteria (Fig. 1a). The selected pose was first evaluated in simulation and then executed on the physical robot (Fig. 1b). CCA selected and executed a complete set of such poses to achieve the required coverage and fidelity, validated the resulting scans, and added each SKU to the cell’s operational representation. Representative final point clouds for two parts from the SKU family are shown in Fig. 1c. CCA commissioned the complete part family in 20 minutes, compared with ~48 hours of manual engineering, and continued to revalidate scan performance at different times of day without engineer involvement.
The second workflow involved removing a worn sanding disc from the robot’s end effector. Heat generated during sanding changes the adhesion between the disc and interface pad, making the required removal interaction variable. Reliable removal requires an approach trajectory and force profile that release the disc without damaging the end effector. Before CCA, an experienced engineer manually taught and validated this trajectory, a process that required approximately half a day.
CCA represented the sandpaper-change station, removal blade, end effector, and relevant interaction regions. The Sensor World Model predicted the force response expected from candidate approach trajectories. CCA selected a promising trajectory, represented by a sequence of trajectory points in simulation (Fig. 2a), and executed the trajectory on the physical cell, as shown at three stages of the interaction in Fig. 2b. CCA evaluated the measured forces and removal outcome and refined the behavior until it satisfied the required success and safety criteria. CCA learned and validated a working removal trajectory in approximately 30 minutes, compared with approximately half a day of manual engineering, without direct human teaching. These results show that CCA can apply a common commissioning mechanism to both part-specific configuration and contact-rich skill acquisition.
