Мобильный робот с манипулятором: как трехуровневая система позиционирования устраняет ошибки при передаче образцов

An autonomous mobile robot can navigate across a laboratory and still miss the sample rack when its arm reaches out. Room-scale navigation brings the platform near a station; it does not guarantee that a gripper, specimen, fixture, and testing machine share the same final coordinates. The last handoff is a separate engineering problem. That distinction matters when a laboratory evaluates a system integrator Singapore project built around a mobile manipulator. The useful architecture is not “AMR plus cobot” as a shopping list. It is a layered positioning stack in which navigation, arm encoders, and visual compensation each remove a different part of the error. Motionwell Automation describes this approach in project P23078, a QA laboratory system using a MiR AMR, Universal Robots cobot, wrist-mounted vision, a 70-position storage rack, and multiple Instron testing stations. The case offers enough public detail to explain how the layers work without claiming a universal accuracy figure. Navigation Solves the Room Not the Handoff An AMR uses its map and sensors to estimate where it is within the laboratory. SLAM-based localization and LiDAR readings help it follow routes, avoid obstacles, and approach a known docking area. That is the right scale for moving between storage, charging, and test stations. Every arrival still contains residual variation. Wheel slip, floor condition, changing landmarks, approach direction, and docking geometry can shift the mobile base slightly. The shift may be small relative to a room, yet important relative to a narrow sample, fixture slot, or machine grip point. Map the Route Before Measuring Final Accuracy Commissioning should separate route performance from handoff performance. First confirm that the AMR can reach the designated station under normal traffic and return to a repeatable docking zone. Record approach paths, obstructions, and failed missions. Only then measure what remains between the docked base and the target. If the AMR cannot reach the zone reliably, adding camera correction at the gripper treats the wrong layer. If it reaches the zone consistently but the handoff varies, the team can focus on arm geometry, calibration, fixture design, and vision. Design the dock as an error boundary. The docking area should define an input window for the next layer. It may use a repeatable approach direction, physical references, known landmarks, or a station pose stored in the map. The goal is not perfect base placement. It is an arrival condition small and predictable enough for the cobot and camera to finish the job. Arm Encoders Narrow the Remaining Position Error After docking, the cobot moves through a taught sequence using its joint encoders. The arm can repeat a pose relative to its own base, but that pose is only useful if the relationship among the AMR, cobot mounting, tool, and station has been calibrated. A shifted mount or changed end-of-arm tool can move the real grip point while the programmed coordinates remain unchanged. The table makes ownership visible. Navigation software cannot correct a loose tool. Cobot repeatability cannot correct a specimen sitting crooked in its slot. Vision cannot rescue a target outside its field of view. Each layer needs an error range the following layer can accept. Calibrate Every Frame That Moves Together The engineering team needs stable transformations among the AMR base, cobot base, tool centre point, camera, and station target. Replacing a gripper, moving the camera, or adjusting the cobot mount can invalidate part of that chain. Calibration status should therefore be treated as maintained configuration, not a one-time setup memory. A practical maintenance procedure identifies which physical changes require recalibration, which reference target to use, who may run the routine, and what result confirms acceptance. Vision Compensation Closes the Final Positioning Gap In the architecture published by Motionwell Automation, a wrist camera detects the sample or target position and converts image coordinates into a correction for the cobot tool frame. This final layer compensates for residual docking offset and target presentation within the calibrated visual range. The camera does not replace the earlier layers. It needs the AMR to arrive close enough that the target appears in the expected search area, and it needs the arm to present the camera from a useful angle. It also depends on controlled target features, lighting, focus, and hand-eye calibration. Keep the Target Inside a Known Search Window A vision routine should have a defined window for position, rotation, and confidence. If the target falls outside that range, the system needs a safe response: retry the view, reposition, prompt an operator, or hold the sample state. Expanding the search window indefinitely can increase false matches and hide a deteriorating dock or fixture. Known-good and deliberately offset samples help establish the boundary. They show whether the system corrects ordinary variation and whether it refuses a condition the gripper cannot safely recover. Verify the Grip Before Moving the Sample Position correction gets the tool to the target; it does not prove that the specimen is secure. Gripper I/O, vacuum or force feedback, part-presence sensing, and a short verification move can provide the next evidence. The correct method depends on the end-of-arm tool and sample. Motionwell Automation describes custom EOAT, gripper control, vision calibration, explicit sample states, and recovery prompts in the laboratory project. Those elements should work as one sequence so a missed pickup does not become a sample that exists in software but remains in the rack. Test the Stack With Deliberate Docking Offsets A clean test from one taught approach can make the stack look more robust than it is. Challenge the system with offsets that remain inside the agreed docking window. Change the approach direction if normal traffic permits it, place a sample at the edge of its allowed seat, and repeat after automatic charging or a long route. A Motionwell Automation review should include ordinary laboratory obstacles as well: a trolley near the route, a station temporarily unavailable, or a rack position held for manual work. The AMR may need to wait or reroute, but the sample state must remain clear while the physical mission changes. Positioning and scheduling meet at that point. Log Which Layer Corrected the Deviation Record the AMR arrival, cobot pose, detected target, applied correction, grip confirmation, and final station result. These values do not need to become a public dashboard, but they help engineers distinguish a navigation drift from a calibration change or fixture problem. The test should also include a condition beyond the allowed range. A trustworthy stack holds or escalates instead of forcing a pickup with low confidence. Recovery behavior is part of precision because it prevents uncertain geometry from becoming an unexplained sample movement. Repeat one recovery with normal laboratory staff present. They should be able to identify whether the hold belongs to navigation, the arm, vision, the gripper, or the test station from the delivered prompts and records. A technically correct stop that nobody can diagnose still creates an operational bottleneck. Measure Three Offsets Before Scaling the Fleet Before adding more routes or robots, measure the errors at the three handoff layers separately. Confirm the AMR reaches a stable docking window, the cobot repeats its calibrated reach, and the camera corrects the remaining target variation without exceeding its search boundary. A mobile laboratory robot becomes dependable when each layer has one job and a visible limit. The route gets the platform near the station, the arm narrows the geometry, and vision closes the final gap. Scaling should follow only after the team can identify which layer failed and recover the sample without losing its recorded state.