Abstract: For the pose error analysis of a polishing robot, on the basis of Vittorio granularity coding method, an improved genetic neural BP network combining binary with real-value coding is proposed with the network architecture and weight vectors optimized. Introducing the binary coding in network architecture to keep the virtues of Vittorio granularity encoding, real-value coding with connecting weighting coefficients, Solis and Wets operation is carried out to bring the virtues of evolutionary programming and evolutionary strategy to the new genetic algorithm. In addition, the combination of genetic algorithm with progeny generated by Solis and Wets operation and optimized individual selection make the genetic search space more diverse to accelerate the genetic algorithm's convergence, with the dynamic parameter coding used instead of Vittorio granularity coding. Optimization of connecting weighting coefficient will be more accurate, and the sharply discontinuous change of adaptability due to Vittorio granularity change can be avoided. Simulation and experimental results indicate that this algorithm can eliminate genetic algorithm's premature convergence and improve effectively the robot pose accuracy.