A practical route to more precise pipetting.

Improve the complete measurement chain: arm, camera, tool, fixture and liquid delivery. Check the resulting error at independent points before trusting a correction.

The current deployment is a working simulation. Camera observations come from rendered images; robot and camera geometry are supplied by the model. No physical accuracy or ISO compliance has been established.

UFactory specifies ±0.1 mm repeatability for the xArm 7. Repeatability describes returning to a pose; it does not establish absolute placement accuracy across the workspace. A longer pipette also magnifies angular error: at a 200 mm lever arm, 0.1° produces approximately 0.35 mm lateral displacement. The engineering conclusion is to observe the tip with feedback and validate locally, rather than interpreting a robot specification as a well-placement guarantee. UFactory specifications.

ROS 2 and MoveIt are in the execution path

The cloud service runs ROS 2 Jazzy, robot_state_publisher, MoveIt 2 move_group and the Pilz industrial planner. The URDF uses the same xArm kinematic chain and collision meshes as MuJoCo. Fixed deck obstacles are transferred to the planning scene. Approved skills request PTP transit and LIN approach/retraction trajectories through GetMotionPlan. MuJoCo executes the returned trajectory through its actuators, checks collisions and actual tip clearance, then verifies the liquid-state transition. Pilz documentation.

Scientist → LLM → schema + protocol approval
  → deterministic skill/state machine
  → localized targets + inverse kinematics
  → ROS / MoveIt planning scene + PTP or LIN
  → MuJoCo actuators + collision/clearance guards
  → deterministic receipts + VLM camera review → UI

Pilz generates industrial motion primitives and rejects colliding paths; it does not search around obstacles. This workcell uses predefined elevated routes. A ROS planning failure stops execution; there is no silent fallback when ROS is configured. The offline installation explicitly identifies its local simulation planner.

Joint feedback is published on /joint_states. The deployed execution adapter is MuJoCo, not a physical ROS controller. A hardware integration must use the official xArm ROS 2 driver, serial-specific kinematic parameters and a checked FollowJointTrajectory execution interface. It must retain stop handling, state checks and calibration provenance. Official xArm ROS 2 packages; MoveIt architecture.

Measure uncertainty on physical hardware

The application uses known simulation geometry and fresh visual feedback. It does not estimate a physical error budget or offer a precision-mode certification. Real-world validation must distinguish repeatability, bias, drift and sensor noise, then propagate measurement uncertainty through the coordinate transforms. Camera reprojection error alone is insufficient evidence of tip-to-well accuracy. NIST: tolerances and uncertainty.

The implemented visual loop requires two independent observations within 0.25 mm, bounded corrections, and a separately checked physical-tip error below 0.5 mm in simulation. Read the actual controller and limits. These thresholds are not physical performance measurements.

What is needed for 50 µm or better?

The following is an engineering roadmap, not an accuracy promise or a list of features already connected to hardware. Define the requirement at the distal tip relative to the well, with load, temperature, speed and confidence level specified.

1. Make the reference stable. Use dowel/kinematic registration, positive plate seating, restrained cables and a short, stiff pipette mount. Survey the nest and board. Warm the arm and optics to equilibrium. Temperature-induced deformation of ordinary metal structures can reach tens of micrometres over laboratory dimensions, so temperature belongs in the record. NIST thermal dimensional metrology.
2. Measure the errors before adding model complexity. Test repeated approaches from identical and opposite joint directions, at cold/warm conditions and with real payloads. Separate repeatability, bias, drift and measurement noise. Monitor changes over time and invalidate calibration after a collision, remount or maintenance event. NIST robot accuracy degradation assessment.
3. Add local closed-loop metrology. Observe both the tip and the well/nest reference near the final pose. Use a calibrated microscope/stereo pair or suitable telecentric optics; command bounded, deterministic corrections only while the tip is safely above the rim. Reject missing features or inconsistent views. Continuous recalibration and visual servoing have been studied specifically for micropipette positioning. Micropipette positioning with continuous recalibration and visual servoing.
4. Use an independent measurement system for very tight targets. NIST demonstrated active tracker corrections on a different industrial robot. Its result depended on the tracker, registration and timing, and does not transfer to this xArm. The useful principle is measured endpoint feedback, including the uncertainty of that feedback. NIST laser-tracker feedback study.
5. Decouple coarse transport from fine placement. If measured xArm repeatability and stiffness consume the error budget, use the arm for transport and a calibrated short-travel XY/Z stage or registered plate stage for the final correction. This is our engineering recommendation, not an implemented hardware feature. Check stage metrology, drift and collision envelopes independently.
6. Fit load and thermal terms only with identifiable data. A 2026 preprint combines geometric, compliance, thermal and gear-transmission effects and reports 26.8 µm mean error on a KUKA KR30, compared with 102.3 µm for its geometric-only model. It supports investigating non-geometric errors, not claiming the same result here. Temporal holdouts and model ablations are especially useful before enabling compensation. Tobuschat et al., 2026 preprint.

Autofocus, a higher-resolution image or an LLM cannot remove unmeasured backlash or tip bending. Keep the LLM in protocol interpretation and the VLM in visible-action review. Deterministic measurement and control must own micrometre-scale corrections.

Position precision and volume precision are separate

Exact simulated conservation does not prove physical dispensing accuracy. Validate the complete pipette-and-tip system gravimetrically at representative volumes. Record water/liquid temperature, density conversion, balance calibration, evaporation correction, environmental conditions, systematic error and coefficient of variation. ISO 8655-6:2022 specifies a gravimetric reference procedure; this application does not claim certification or reproduce the complete standard. ISO 8655-6:2022.

Use liquid-specific aspiration/dispense speeds, dwell, immersion depth and pre-wetting. Viscous or foaming liquids may need reverse pipetting or positive displacement; volatile liquids need attention to pre-wetting and evaporation. These methods affect residual/waste volume and must be represented in the protocol and liquid ledger before being enabled. They are hardware commissioning recommendations, not hidden corrections to current simulated volumes. Eppendorf technical guide on pipetting accuracy.

Never use an overhead VLM to certify an exact volume. Use piston/pressure feedback for operation checks and calibrated mass or other appropriate metrology for accuracy validation.

Hardware acceptance before autonomous use

1. Install the official arm driver and serial-specific model. Confirm frame conventions, joint order/limits, tool payload and emergency-stop behavior before powered workcell motion.
2. Calibrate camera intrinsics and independent optical/tracker references. Lock focus and mounts; record artifact dimensions and uncertainty.
3. Measure tip-to-target errors over the intended workspace. Reserve independent observations and a different-time repeat dataset for acceptance.
4. Validate tool-to-well errors at plate corners, interior wells, both approach directions and different tip lots. Measure radial, vertical and angular error; report maximum, RMS, bias and repeatability separately.
5. Repeat at realistic thermal and load conditions. Reject out-of-domain poses. Record per-tip runout and test intentionally mis-seated/occluded tips.
6. Check failed localization, lost ROS service, stop during motion, camera disconnection and uncertain VLM review. No silent continuation or automatic repeat of an ambiguous transfer.
7. Perform liquid-volume validation independently. Only claim a physical accuracy specification after independent measurements support its total error budget.

The current simulator’s extra ground-truth guards are independent verification aids. On physical hardware they must be replaced by measured interlocks, surveyed geometry and validated sensors—not treated as available information.

References used by this design

The current implementation uses model geometry, image-based station localization, bounded position-based visual feedback and independent simulation guards. It contains no calibration estimator or precision-mode profile.

• Chaumette & Hutchinson: Visual Servo Control, Part I — background for position-based feedback.
• OpenCV rigid marker boards — jointly using multiple tags from one fixed fixture.
• MoveIt/Pilz motion primitives — the deployed PTP/LIN planner.
• NIST: tolerances and uncertainty — physical validation must include the sensing chain.

Implementation: perception.py estimates fixture poses; visual_servo.py measures the tip and bounds corrections; robotics.py validates returned trajectories; ros_adapter/ runs MoveIt services. Experiment observations and images remain in the deployment data volume for audit.