Bin picking fails when a robot treats uncertainty as a fault. Parts settle in unknown orientations, so the position estimate may miss tolerance, aborting the grasp and stopping the line. That response can work when parts and fixturing are consistent and uncertain cases are rare. In bin picking, uncertainty is routine. The planner must decide what to do when the part location is unclear.
Random orientations, occlusion, changing light, and noisy depth returns on specular surfaces make uncertainty routine. Reliable picking therefore needs planning that treats uncertainty as an input, not a failure condition.
Detected Uncertainty Versus Acted On Uncertainty
Picking systems often compute a pose confidence score, compare it with a threshold, then proceed or fault. That misses useful distinctions. A score of 0.55 for one grasp and 0.89 for another both clear 0.5, but they call for different actions. An 89% grasp on an exposed part can proceed. A 55% grasp on an occluded part should reposition the camera or use a compliant approach.
Neuralzome carries uncertainty into planning. Perception returns a belief distribution over possible part states, not one pose with a binary flag. The policy engine selects actions from that full distribution rather than a point estimate.
Belief in a Real Time Control Loop
Full belief state planning is computationally costly. Exact Bayesian inference over a continuous pose distribution will not fit a 50ms control loop. We factor uncertainty across the axes most relevant to grasp success: position in the x and y plane, rotation about the vertical axis, and tilt from horizontal. These three degrees of freedom cover most failures in our studied bin picking cases.
A particle filter maintains the belief with a small count in normal operation and expands it when state entropy crosses a threshold. It updates on each depth frame, synchronized with the 10ms sensor cycle. Policy inference reads the filter at each planning tick. On commodity embedded hardware, the pipeline fits within the 50ms ceiling, leaving time for action selection.
This is not theoretically optimal. A full factor graph would provide more principled uncertainty estimates. The practical gain is running on available manufacturer hardware at production control latency.
Action Selection Under Uncertainty
Given a belief distribution over part state, the policy engine chooses the action with the highest expected task success: take the best grasp, reposition the wrist camera or apply a structured light pulse, or use a compliant approach that tolerates pose error.
Sensor actions add latency, typically 150 to 400ms per maneuver, and that cost compounds in high throughput cells. The policy compares expected information gain, based on current entropy and historical disambiguation rate, with expected grasp improvement and cycle time.
In pilot runs across 2024 and 2025, sensor actions occurred on roughly 12% of picks. Their later grasp success matched picks with high initial confidence. The other 88% proceeded directly and performed well because the threshold selected genuinely high confidence cases. Too conservative a threshold adds unnecessary sensor moves; too aggressive a threshold spends attempts on weak estimates.
Occlusion as a Structured Problem
Occlusion remains the hardest bin picking case, and it is not solved generally. We make the response systematic. When the top accessible part is heavily occluded, perception returns high tilt entropy. The policy checks for lower entropy candidates and routes to them. If the accessible layer is heavily occluded, it can trigger bin agitation when supported or escalate to a human in the loop handoff.
The system knows why it is uncertain. High entropy in tilt axis differs from high entropy in x-y position and needs different recovery. This is uncertainty flowing into planning. A binary confidence flag loses that structure; a belief distribution keeps it.
What This Does Not Fix
Uncertainty aware planning does not solve out of distribution cases. If an unseen part arrives, perception returns high entropy and planning responds conservatively or escalates. This does not mean out of distribution coverage is solved. It means the response is appropriate while policy coverage remains bounded by training data. Expanding it requires demonstrations and policy updates, not only inference changes.
Representable uncertainty is not always task relevant. We focus on grasp related pose uncertainty because it drives bin picking failures. Tasks centered on material properties, part weight, or internal geometry need another representation. We tell customers where this architecture fits and where another approach is needed.
Practical Results
The clearest data came from a plastics components facility handling roughly 200 SKU variants, from flat rectangular clips to irregular housings. The uncertainty aware architecture did not remove failures. It made them predictable and recoverable. The system escalated reliably when uncertain and proceeded successfully when confidence cleared the line efficiency threshold. Unrecoverable jams fell significantly versus the prior system, which lacked uncertainty representation. These were internal pilot benchmarks, not a controlled study, but matched the operations team's reported shift.