Abstract: With the popularization of robot technology in orchard agricultural production, orchard robot has been applied more and more in actual production. Among a number of technologies supporting orchard robots to achieve automation, real-time local trajectory optimization is an important guarantee for the safe and stable autonomous operation of orchard robots. Aiming at the problems of the original Augmented Lagrangian TRajectory Optimizer (ALTRO) algorithm, such as low iteration efficiency, easy to fall into numerical sickness and difficult weight balance, This paper presents a real-time local trajectory optimization algorithm for orchard robot based on improved ALTRO. On the premise of obtaining the input of the robot's global reference trajectory, the algorithm firstly adopts the accelerated augmenting Lagrange algorithm to replace the augmented Lagrange link of the original algorithm, improves the multiplier iteration strategy of ALTRO algorithm, and iteratively computes Lagrangian multipliers with higher convergence rate and more stability to achieve accelerated convergence of the algorithm. Secondly, due to the introduction of the Lagrange link in the algorithm, in order to ensure the numerical stability of the algorithm, it is necessary to ensure that the multiplier obtained by each iteration update is within the scope of the feasible domain. Therefore, by adding the multiplier feasible domain projection method, the updated multiplier value is projected onto the feasible domain space to ensure the rationality of the multiplier update. The numerical ill-condition of the algorithm caused by many iterations is fully avoided, and the numerical stability of the algorithm is improved. Finally, in view of the difficulty of balancing multiple cost weights in the original ALTRO algorithm, the adaptive scaling factor based on trajectory time step is introduced to dynamically adjust the weight of the distance from the end point in the algorithm, which avoids excessive attention to the pose of the target end point and enables the algorithm to respond better to local obstacles in the trajectory optimization process. Experimental results show that, given the same global reference path and basic parameter configuration of the algorithm, compared with the original ALTRO algorithm, the improved ALTRO algorithm proposed in this paper has a computation time reduction rate ranging from 32.76% to 67.80%. In addition, the reference trajectory obtained by the optimization of the algorithm in this paper reduces the discontinuity and mutation of the state and control variables in the trajectory optimization process, and the change rate of the robot state variables and control variables included in it is more stable, so that the trajectory optimized by the algorithm in this paper performs better than the original algorithm in terms of geometric smoothness and dynamic smoothness in the comprehensive evaluation. Thus, it can provide a better reference for the robot's running trajectory.