Assembly order stops needing after-the-fact repair: a constraint-preserving planner returns the exact best remaining sequence in all 128 recoverable disturbance cases
Synopsis
In modular assembly the robot's task order must respect assembly precedence and be reordered when conditions change after work begins. Constraint-Aware Genetic Task Planning (CGTP) keeps initialization and variation inside the precedence-feasible region and, on replanning, freezes completed operations and re-optimizes only the remaining task graph. On an exhaustively verifiable eight-task benchmark it returned the exact best feasible suffix in all 128 structurally recoverable scenarios and flagged all 3 unrecoverable ones; on a 90-instance corpus, per-class recovery ran from 0.881 to 1.000 with no budget-conditional search failures.
Interpretation
An assembly planner can now search over task orders without breaking assembly precedence, and after a mid-execution disturbance it reorders only the operations not yet done. Conventional genetic algorithms vary candidate orders with ordinary crossover and mutation, then rely on a repair step to pull illegal orders back into the feasible region; CGTP instead uses operators that only produce precedence-satisfying orders. On a nine-cell corpus (task counts 12, 16 and 30 crossed with sparse, moderate and dense constraint regimes, ten instances per cell and fifty paired trials each), CGTP triggered no precedence repair at any point in the campaign.
Replanning after a disturbance produced the exact answer at a scale small enough to check by enumeration, not merely a feasible one. Evidence checking replanner output case by case at an exhaustively enumerable scale was previously lacking. Of 168 disturbance scenarios on the eight-task benchmark, 131 were geometrically admissible; on all 128 structurally recoverable ones the replanner returned the best feasible suffix confirmed by exhaustive enumeration, and all 3 structurally unrecoverable cases were correctly flagged.
At corpus scale recovery is high, and the unrecovered remainder is attributed to structural infeasibility rather than to the planner failing to find a solution. Earlier accounting lumped non-recovery into a single failure rate, unable to separate an unsolvable problem from an insufficient search budget. Across 67,500 replanning runs on 90 instances, per-class recovery was about 0.978 for D1, 1.000 for D2 through D4 and about 0.881 for D5, with zero budget-conditional search failures; the residual was certified structurally unrecoverable by an independent oracle.
On static objective quality CGTP does not beat repair-based GA, and this trade-off shifts with problem size. Repair-based GA is better at the largest size tested. The paired CGTP-versus-repair-GA difference was statistically indistinguishable at N of 12 and 16, while repair-based GA was significantly better in all three N=30 regime cells (+0.029 sparse, +0.023 moderate, +0.007 dense, Holm-adjusted p=0.018).
Perspective
The results cover task-level precedence planning and disturbance replanning: exact verification on the eight-task benchmark, recovery statistics on the 90-instance corpus, and a pose-level audit in a simulated Franka Emika Panda workcell (96/96 audited poses inverse-kinematics reachable and collision-free). Full motion-level executability, closed-loop execution feedback and large-scale optimality are outside what is demonstrated.
The corpus has only ten instances per cell, which bounds cell-level statistical power; static comparisons use a single frozen budget and budget sensitivity is not characterized; disturbances are injected at one point only; corpus-scale recovery evaluates CGTP alone, so it is a capability and attribution result rather than a comparison with another replanner; and for D5 in the N=12 and N=16 dense cells the preregistered tests did not reject the 0.5 recovery-floor null at the corrected threshold.