Human-AI symbiosis in industrial automation should not be treated as ageneral argument for keeping humans in every automated process. The morefundamental design question is where collaboration between humans and AI isstructurally needed, and where it can be superior to both full automation andmanual work alone. This paper addresses this question through the case ofrobotic recycling. Post-consumer plastic sorting is often approached with thegoal of automating as much as possible, because manual sorting is physicallydemanding, unpleasant, and difficult to scale. However, participatory fieldworkin lightweight packaging sorting facilities shows that sorting work is not astable, repetitive pick-and-place task. Workers continuously adapt toheterogeneous material streams, changing waste sources, weather-dependent anddamaged object conditions, conveyor dynamics, disruptions, and varying economicvalue. The resulting industrial allocation pattern is often binary: selectedsubtasks are automated as far as possible, while the remaining work is left tohumans as residual manual work, a pattern previously discussed as"left-over allocation''.Building on this diagnosis, we argue that robotic recycling is a genuinehuman-AI symbiosis problem. A recycling robot must commit to a pick at aspecific moment in time, although material identity, graspability,contamination risk, downstream value, and future waste-stream composition areonly partially known. Human operators, meanwhile, possess tacit and situatedknowledge about which objects are worth picking, which objects are risky, andwhen the sorting strategy should change. We therefore propose anRLHF-inspired operator-prior architecture for RecycleBot. The approachformulates robotic sorting as a value-aware decision problemand combines a learned action-value model with a human-derived behavioralprior. The resulting KL-regularized policy treats human expertise not as afallback after automation fails, but as a structured prior over robot actions.A rolling assignment layer extends the formulation to short-horizon planningunder robot-capacity constraints. Synthetic multi-seed experiments show thatthe operator-prior policy performs best of all compared policies early in arun, avoids the run-to-run failures of unregularized value learning, andreduces the number of decisions requiring human correction by a factor offive to seven, at a cost in long-run reward that the KL temperature controlsdirectly. The contribution is analgorithmic and architectural pathway from automation-as-replacement towardadaptive human-robot collaboration in industrial recycling which can be adapted to other domains with high variability.