Recovery Policy Optimization for Online Quality Monitoring and Hybrid Force/Position Control in Robotic Precision Assembly

Contact-rich precision assembly fails in ways visible in the force/torque and proprioceptive streams before they are visible in the finished part, which makes online quality monitoring attractive. We ask whether monitoring accuracy or the recovery policy is the binding constraint on closed-loop performance, and answer it by improving both. For monitoring we propose AdaStack, which mixes a random forest on summary statistics, a fully convolutional network and a CNN-LSTM with simplex weights fitted on pooled out-of-fold posteriors, so the combination adapts to whichever learner suits a given task. On the UCI Robot Execution Failures benchmark, evaluated over repeated stratified cross-validation against eight baselines under one protocol, AdaStack attains the best mean rank (3.20 against 4.00) and the best mean macro F1 (0.787 against 0.784), and is the top model on three of five tasks; it significantly outperforms dynamic time warping on two tasks after Holm correction, while differences against the stronger baselines remain within the resolution of a 47-to-164-trial benchmark. We then close the loop in simulation, on a physics-based MuJoCo peg-in-hole cell. There the limiting factor turns out not to be perception: with a recovery policy written by hand, a spiral search that ignores the estimated state seats 78.5% of insertions while the learned monitor reaches 33.0% and even an oracle given the true fault state reaches 52.0%. Optimizing the twelve free policy parameters against a cost over seating, contact force, cycle time and damage, with disjoint optimization and evaluation seeds, yields in this simulated cell a controller that matches the search’s seating rate (76.5% against 78.5%) while cutting mean peak contact force by 46% (17.2 N against 31.6 N) using the learned monitor rather than an oracle. The same gain, however, is available without the monitor: a spiral search whose own three parameters are optimized with the same budget seats 79.0% at 18.3 N, and it matches or beats the monitor-driven policy under four cost weightings and five of six shifted operating conditions. What paid was optimizing the reaction, not conditioning it on the estimated state. Effort spent on the recovery policy repaid far more than effort spent on monitoring architecture; whether this transfers to a physical cell remains to be tested.