When we started building Neuralzome, we ran our first external pilots in relatively controlled environments: single-product assembly cells with consistent part geometry, stable lighting, and cooperative floor teams. We got useful signal from those deployments, but we were not really testing the hardest version of the problem.
The hardest case is a high-mix factory: one cell builds four or five product families, changes between them several times per shift, and treats the robot as one fixture among many in a tight layout. We spent most pilot time through mid-2025 in this setting. It shaped the Neuralzome architecture more than any benchmark dataset.
High-Mix at the Robot Level
High-mix manufacturing means one robot arm assembles products with different geometries, fixture configurations, and sometimes process sequences. Volume per variant is too low for a dedicated robot, so the same arm must handle every variant.
In our pilot, a cobot arm handled four fixture variants for connector housing assembly. Variants A and B had housing widths differing by 8mm and needed different gripper approach angles. Variant C added a snap-fit latch requiring a brief downward press at insertion's end. Variant D used another pallet format with greater positional variance during loading.
A traditional teach-pendant system needs four waypoint sets and four programs. At changeover, the operator selects one. The robot follows that sequence exactly, until a part is misplaced, the fixture shifts, or a new pallet's tolerances differ from the programmer's nominal specification.
Why Single-Variant Policies Failed to Transfer
We first trained one policy per variant and switched policies on a changeover signal. It worked, but not instantly. Each policy saw demonstrations from only its own variant, including visual features tied closely to that fixture's geometry. Policy A used variant A's fixture edges to localize grasps. On variant B's slightly different edge profile, its confidence dropped, and it sometimes requested more sensor readings before picking. The issue was not the signal itself. The policies had learned variant-specific visual evidence.
This was not catastrophic, and the robot still operated correctly. Yet the hesitation showed dependence on visual features that did not generalize. Adding a fifth variant would mean a fifth policy, with each becoming more fragile as the new visuals diverged from training data.
We needed policies aware of variants without overfitting to them.
Multi-Variant Policies
We shifted to joint training on demonstrations from all variants, adding the variant identifier to the policy context. The variant ID is not a class label for selecting a mode. It is a continuous context vector that shifts the output distribution. The policy learns which visual-force combinations matter across variants for grasping, and which features are specific and should receive different weight by context.
This also changed our demonstration protocol. We had first recorded each variant in its own session. Multi-variant training required interleaved demonstrations, preventing temporal patterns in the data from silently identifying the active variant.
Interleaving exposed another issue. Floor technicians, rather than robotics engineers, naturally hesitated and made micro-adjustments when changing variants mid-session. The recordings captured longer dwell times near the fixture before a pick. The policy treated that transition behavior as valid, making changeover slightly slower than steady-state operation learned within one variant.
We removed the first two or three picks after each variant switch from the demonstrations. The resulting policy handled changeover more cleanly and matched single-variant cycle time during steady-state runs.
Fixture Drift
Across six months, policies approved at a production run's start sometimes degraded during the final hour of an eight-hour shift. Fixtures were drifting from thermal expansion and vibration. In some cases, position moved 4-6mm from nominal by shift end.
Grasp localization combined visual pose estimation with prior fixture knowledge. For some orientations, 4mm fell within visual uncertainty, so the policy missed the shift at first. It kept picking farther off-center until drift passed the grasp success envelope and caused failure.
We made two changes. A continuous fixture position estimator now tracks observed position across picks and maintains drift from nominal. Above 3mm, it signals a switch from nominal-fixture-based to observed-fixture-based localization. We also added a periodic re-reference check: every N picks, typically 50, the robot performs a short scan-and-estimate routine to refresh fixture state instead of relying on accumulated drift.
Neither change existed in the original architecture. Both came from production failures that a lab demonstration would not reveal.
What Floor Complexity Taught Us
We underestimated the production floor's distance from our controlled prototypes. Three surprises were especially important.
Lighting varied more than expected. Natural and fluorescent sources mixed, while cloud cover and open or closed factory doors changed their ratio. The depth camera adapted exposure, but the RGB stream used for visual features did not adapt quickly enough, lowering some pick confidence during transitions. We added an ambient-light monitor to flag variable periods and adjust confidence thresholds.
Other cell machines introduced vibration that shifted the force/torque sensor baseline. Wrist-sensor noise increased while a nearby press cycled, a condition absent from our lab characterization. We added vibration-aware baseline calibration to re-zero the F/T sensor at each pick cycle instead of using only startup calibration.
Production batches varied more than the nominal drawing suggested. A critical dimension specified +/- 0.3mm, while some batches reached about +/- 0.8mm on one axis. Demonstrations came from one batch, so the policy learned that batch's variation rather than the full population. We expanded demonstrations across multiple batches before it generalized correctly.
What We Would Change
Some lessons require deployment. Fixture drift, batch variation, and lighting dynamics can only be characterized in real production. Earlier, we could have added explicit production-drift monitoring to policy evaluation, letting degradation surface sooner and trigger re-characterization instead of quietly accumulating until failure.
We are building that capability now. Production policy health monitoring will flag sessions that depart from expected confidence distributions. The priority comes from repeated quiet degradation followed by threshold crossing, not from any catastrophic deployment failure.