In recent years, methods based on deep reinforcement learning (DRL) have received significant attention in the field of picking path planning. However, existing research is primarily focused on the picking manipulator, while the close coordination between the mobile chassis and the manipulator is often overlooked. As a result, the full potential of the robot in terms of mobility and manipulation is not fully utilized. To address this limitation, a collaborative path planning method for litchi picking robot was proposed using an improved twin delayed deep deterministic policy gradient (TD3) algorithm. Initially, the collaborative path planning problem was modeled using deep reinforcement learning. The state and action spaces were defined according to the requirements of the task, and a composite reward function was formulated, taking into account position, orientation, and obstacle avoidance. Subsequently, the TD3 algorithm was enhanced through the integration of dual Actor networks and layer normalization. Finally, an unstructured litchi picking environment was developed on the CoppeliaSim simulation platform, and the robot was trained within this simulated environment. The simulation results show that deep reinforcement learning effectively addresses the collaborative path planning problem. The improved TD3 algorithm achieves faster convergence. In the testing phase, the path planning success rate reaches 96%, representing an improvement of nearly 69.9 percentage points over the original algorithm. These results indicate that the proposed system effectively overcomes the challenges posed by high-dimensional state-action spaces in collaborative path planning. It enables coordinated motion between the mobile chassis and the manipulator, thereby enhancing the robot’s operational intelligence. Furthermore, it demonstrates a clear advantage in planning success rate and offers valuable guidance for future research in picking robot path planning.