An Improved DDPG-based Trajectory-Velocity Planning Method for Compliant Collection of Cluster-type Fresh Fruits

• Improved DDPG for synergistic trajectory-velocity planning in fruit packing. • Human demonstrations enhance RL efficiency for stochastic cluster handling. • RL-based velocity regulation ensures compliant handling during fruit packaging. • The method reduces contact impact and improves damage-free collection success. Fresh fruit packaging still relies heavily on human expertise, while existing robotic methods are primarily designed for efficiency and often pay insufficient attention to the compliant handling required for delicate fruits. To address the demand for compliant manipulation during the automated collection and packaging of bunch-shaped fresh fruits, this paper proposes a robotic arm trajectory-velocity planning method that integrates human operational skills. Based on the characteristics of the fruit collection and packaging tasks, motion constraints are established to prevent collision damage to fruit clusters with varying lengths. Human-demonstrated motion trajectories for fruit placement were collected as experience replay. By incorporating a Deep Deterministic Policy Gradient (DDPG) reinforcement learning model, a reward mechanism designed for the synergistic optimization of 3D trajectories and velocities was proposed. Consequently, a compliant motion planning model for the collection of fruit clusters with stochastic morphologies was established. Compared with classical reinforcement learning models such as SAC and TD3, the proposed method achieves synergistic optimization of the robotic arm's trajectory and velocity, with an average endpoint error of 0.0182m and an average midpoint error of 0.0219m. Furthermore, relative to the baseline DDPG model, the success rate of collision-free collection is improved by 16%, and the instantaneous contact velocity between the fruit cluster and the packaging container is reduced by 22.52%. This research provides a methodological reference for developing robotic packaging systems, with its effectiveness further confirmed through performance evaluation on a physical robotic platform.