Abstract: High-density dwarf culture has the characters of high yield and easy mechanization, thereby to be expected as a new development direction of the fresh fruit industry. In harvesting, the mechanism of multi-arm cooperation can remarkably improve the operation efficiency of agricultural robots in the orchards of high-density dwarf culture. Generally, multiple picking zone is are established for the collaboration of multi-arm harvesting robots, in order to assign proper picking objects to every manipulator. However, these different zones can be partially overlapped, leading to potential conflicts of manipulators in a robot. Path planning can be utilized to generate a trajectory leading the tip of the manipulator to the goal without collisions. In this study, a global optimization using a Genetic Algorithm(GA) was proposed to solve the planning problem of harvesting tasks with overlapped picking domains in the multi-arm robots toward high-density dwarf orchards. 1) A multi-arm structure without a dead zone in the fruit harvest was specially formulated as a sort of asynchronously multiple traveling salesman problem with overlapped zones. The following mild assumptions were made: a) The visual sensing system can precisely locate the fruits to be harvested; b) The fruits covered by leaves were not considered; c) The exceptional cases during the switching phases from one fruit to another were not considered in the planning stage, meaning that the traveling time linearly depended on the Euclidean distance. Five rules were established for the manipulators to work with better cooperative behaviors in a safe region of the workspace, where no collision occurred. 2)A modified GA was applied to optimize the collision-free trajectory planning of a flexible manipulator, further to ensure operational safety without conflicts in the shortest traveling time. Gene codes were used to formulate various domain-manipulator pairs, thereby determining picking sequences in a manipulator. The total traveling time of travelers was the objective to be optimized, where the population of gene codes was optimized iteratively in the proposed GA. In this case, the critical phases were coding and decoding. The double chromosomes were selected to formulate various picking sequences. After that, three operators were introduced, including selection, crossover, and mutation. The rule of synchronization was designed to avoid different travelers visiting the same cities simultaneously, to cope with the asynchronism in a queuing situation.3) For the three fruit distribution scheme, Once the optimized solution was obtained, the planner can easily achieve the proper sequences for each manipulator to pick fruits via decoding the double chromosomes. As such, each manipulator performed the corresponding task, and the overall picking time was therefore reduced. The test results showedthat:1) The proposed planning converged at 500 and 2 000 iterations, when solving 43 fruits and 90 fruits planning problems with 4 manipulators.2) Compared with the sequential traverse, the modified GA increased the efficiency by 40.9% and 54.98%, when harvesting 43 fruits and 90 fruits, whereas, increased the efficiency by 4.25 times, compared with a single manipulator robot when harvesting 90 fruits.3) compared with the sequential traverse and random traverse, the modified GA increased the efficiency by 110.69% and 27.18%, 20.45% and 23.33%, 12.94%, and 21.69% under three different distributions, respectively. The contributions can be: a) A new planning strategy was proposed for the multi-arm harvester to avoid collisions, where each manipulator behaved cooperatively. b) A novel genetic algorithm was presented, where coding and decoding were first proposed in this field. c) The rule of synchronization was first proposed to deal with the simultaneous visits in the queuing phenomenon during planning. d)A four-arm cooperative harvester was used to verify the effectiveness of the system, where the ergodic picking of each arm can be achieved without conflicts within a minimum duration. The framework can be generalized to many configurations of harvesters, suitable for robots with different numbers of arms, different varieties of fruit, and different ranges of overlapped domains.