Obstacle-Avoidance Path Planning for a Basket-Carrying Robotic Arm Based on an Improved APF-RRT Algorithm

Technical Summary: Obstacle-Avoidance Path Planning for a Basket-Carrying Robotic Arm Based on an Improved APF-RRT Algorithm

Problem Statement
The paper addresses the critical challenge of path planning for material-handling robotic arms in unstructured agricultural environments. Specifically, it targets the task of autonomously grasping and transporting field baskets amidst complex, randomly distributed obstacles (e.g., crop branches) and highly variable basket poses. Conventional Rapidly-exploring Random Tree (RRT) algorithms struggle in these high-dimensional (6-DOF) joint spaces due to weak goal-directedness, low convergence efficiency, and the generation of redundant, non-smooth paths. Furthermore, standard Artificial Potential Field (APF) methods suffer from local minima traps and goal-unreachability issues when targets are near obstacles. These limitations hinder the real-time responsiveness and motion stability required for safe agricultural conveying operations.

Methodology
The authors propose an improved RRT algorithm that deeply integrates a modified APF method into the node-expansion mechanism. The core technical components include:

1. Multi-Field Coupled Node Expansion: Instead of relying solely on random sampling, the algorithm computes a resultant force vector for node expansion by combining three components:
   • Goal Attraction: A unit vector weighted by an attraction gain coefficient.
   • Obstacle Repulsion: Based on the gradient of a repulsive potential field.
   • Random Exploration: A unit vector toward a random sampling point to maintain global coverage.
2. Power-Law Decay Repulsive Gain: To solve the goal-unreachability problem (where repulsion prevents the arm from reaching a target near an obstacle), the authors replace the standard exponential decay with a power-law decay model for the repulsive gain coefficient ($k_r$). This ensures the repulsive force diminishes at a higher-order rate ($O(d_{goal}^\beta)$) as the node approaches the goal, allowing convergence in narrow spaces.
3. Three-Level Progressive Escape Mechanism: To address local minima where attraction and repulsion forces cancel out, the algorithm employs a failure counter ($f$) and a three-tiered escape strategy:
   • Normal Mode: Standard expansion.
   • Soft Escape: Adds Gaussian-distributed random perturbations to the expansion direction when minor stagnation is detected.
   • Hard Escape: When deep stagnation occurs, the algorithm suppresses potential-field guidance, generates multiple candidate directions, and selects the one with the greatest collision-free traversal span.
4. Penalty-Weighted Distance Metric: To prevent the nearest-neighbor search from repeatedly selecting nodes trapped in local minima, a penalty term based on the cumulative failure count is added to the Euclidean distance metric.
5. Adaptive Strategies:
   • Adaptive Step Size: Dynamically adjusts the expansion step size based on proximity to obstacles and the goal, balancing exploration speed with collision safety.
   • Dynamic Goal-Bias Probability: Adjusts the probability of sampling the goal configuration based on iterative progress, using a piecewise smooth function (cubic Hermite interpolation followed by linear growth) to transition from broad exploration to directed convergence.
6. Post-Processing: The initial path undergoes multi-pass greedy pruning to remove redundant nodes, followed by cubic B-spline smoothing to ensure $C^2$ continuity (continuous curvature and angular acceleration) for smooth robotic motion.

Key Contributions

• Algorithmic Integration: A novel fusion of improved APF and RRT specifically tailored for the high-dimensional, obstacle-dense joint spaces of agricultural robotic arms.
• Goal-Unreachability Solution: The introduction of a power-law decay repulsive gain coefficient effectively eliminates the goal-unreachability dilemma common in traditional APF methods.
• Active Local Minima Escape: The development of a three-level progressive escape mechanism combined with a penalty-weighted distance metric allows the random tree to actively bypass deadlock regions rather than getting stuck.
• Engineering-Oriented Optimization: The inclusion of adaptive step-size and dynamic goal-bias strategies, alongside rigorous post-processing (pruning and B-spline smoothing), ensures the generated paths are not only collision-free but also computationally efficient and kinematically smooth.

Experimental Results
The algorithm was validated through comparative simulations in 2D and 3D configuration spaces and real-robot experiments on a 6-DOF Siasun GCR25-1800 collaborative robotic arm.

• 2D Simulations: Compared to conventional RRT, the improved algorithm reduced average iteration counts by 85.8%–89.7% and planning time by 42.1%–56.4% across sparse, narrow-corridor, and complex-dense obstacle environments. Path lengths were reduced by 18.6%–19.6% compared to RRT and RRT-Connect.
• 3D Simulations: In a 3D space with spherical obstacles, the algorithm completed planning in an average of 17.1 iterations (a 99.3% reduction vs. RRT and RRT*) and shortened path length by 31.4% compared to RRT.
• 6-DOF Robotic Arm Experiments: In joint-space planning for basket carrying, the proposed algorithm reduced path cost by 38.9% (vs. RRT), 43.9% (vs. RRT-Connect), and 32.9% (vs. RRT*). Planning time was reduced by 10.1%–20.9% compared to the baselines.
• Real-Robot Validation: Physical trials confirmed that the planned trajectories resulted in smooth, jitter-free motion across all joints with no collisions, validating the algorithm's stability and engineering applicability.

Significance and Claims
The paper claims that the proposed algorithm effectively satisfies the engineering requirements for planning responsiveness and motion stability in complex agricultural environments. By addressing the specific limitations of standard RRT (inefficiency, redundancy) and APF (local minima, goal unreachability), the method provides a robust solution for automated material handling in unstructured farmland settings. The authors note that while the algorithm performs well in semi-structured, relatively static scenarios (as tested in the lab), it currently relies on pre-modeled static obstacles and does not yet incorporate online re-planning for dynamic, unpredictable field conditions (e.g., swaying branches or moving personnel). The study positions the work as a significant step toward practical agricultural automation, offering a feasible algorithmic foundation for robotic arm conveying operations.