*Citrus-Picking Robot Arm Achieves Faster, Smarter Path Planning with Enhanced Informed-RRT Algorithm**
In the race to automate agriculture, one of the toughest challenges has always been fruit harvesting—especially when it comes to delicate crops like citrus. Unlike structured factory floors where robotic arms operate with predictable precision, orchards are chaotic, unstructured environments filled with unpredictable obstacles: branches, leaves, uneven terrain, and fruits that grow in random orientations. For a robot to pluck a single orange without bruising it or knocking over its neighbors, it needs more than just dexterity—it needs intelligence in motion.
Enter a breakthrough from researchers at Chongqing University of Technology: an improved version of the Informed-RRT* path planning algorithm that dramatically boosts the speed, success rate, and efficiency of citrus-picking robotic arms. Their innovation doesn’t just tweak an existing method—it rethinks how robots explore complex spaces by introducing a “pre-picking guide point” and running four search trees in parallel. The result? A system that plans paths nearly twice as fast, succeeds 11% more often, and delivers smoother, shorter trajectories—all critical for real-world agricultural robotics.
This isn’t theoretical tinkering. The team validated their approach in both 2D simulations and full 3D ROS-based robotic environments, using a real six-degree-of-freedom collaborative arm (the Aubo-i5) to mimic actual field conditions. And the numbers speak loudly: average planning time dropped to just 0.81 seconds, with a 96% success rate across diverse simulated orchard layouts. Compared to the original Informed-RRT*, their enhanced algorithm cut planning time by 46%, reduced path cost by 16%, and slashed node count and iteration cycles by over 40%.
For an industry desperate to address labor shortages, rising costs, and the need for sustainable farming, this work represents a meaningful leap toward practical, deployable harvesting robots.
The core problem in robotic fruit picking isn’t grasping—it’s getting there. Most existing systems rely on classical sampling-based planners like RRT (Rapidly-exploring Random Tree) or its optimized cousin RRT*. These algorithms work by randomly probing the robot’s configuration space—essentially its possible joint angles and positions—to find a collision-free route from start to goal. They’re powerful in high-dimensional spaces, which is why they’re widely used in robotics. But they have well-known flaws: slow convergence, suboptimal paths, and inefficiency in cluttered environments.
Informed-RRT*, introduced in 2014, tried to fix this by focusing future sampling only within an “informed set”—a hyper-ellipsoid defined by the current best path length and the straight-line distance between start and goal. This narrowed the search area, speeding up optimization. But it still suffered from a critical dependency: the quality of the initial path. If the first feasible route was long or winding, the ellipsoid became bloated, diluting the benefits of focused sampling.
Worse, in agricultural settings, the end goal isn’t just reaching a point in space—it’s approaching a fruit from a specific angle, usually horizontally, to avoid damaging stems or neighboring produce. That often requires a two-stage maneuver: first move to a staging position near the fruit, then adjust posture for the final grasp. Traditional planners handle this by running two separate pathfinding sessions—one to the intermediate point, another to the target. This doubles computation time and increases failure risk, especially when the environment changes between attempts.
The Chongqing team tackled this head-on. Instead of treating the pre-pick position as an afterthought, they embedded it directly into the planning architecture. They defined a fixed “pre-picking guide point” located 25 centimeters horizontally from the target fruit—a distance empirically chosen to allow safe, stable approach while minimizing unnecessary motion. Then, rather than running two sequential searches, they launched four randomized trees simultaneously: one from the robot’s starting pose toward the guide point, another from the guide point back toward the start, a third from the guide point toward the fruit, and a fourth from the fruit back toward the guide point.
This four-tree parallel expansion creates a much denser exploration front in the critical zones of the workspace. It’s like sending out four scouts instead of one, each covering overlapping but complementary ground. When any two opposing trees connect—say, the start-to-guide and guide-to-start trees merge—they form a complete segment. Do this twice (start→guide and guide→goal), and you get a full, continuous path in a single planning pass.
But parallelism alone isn’t enough. Random sampling is still wasteful. So the researchers added a probabilistic bias: at each expansion step, there’s a small chance (set to 10% in experiments) that the next sample is pulled directly toward the target or guide point, not a random location. This “goal-biased sampling” accelerates convergence without sacrificing completeness—the algorithm remains probabilistically complete, meaning it will eventually find a solution if one exists.
They also introduced a dynamic step size mechanism. Traditional RRT variants use a fixed step length when extending branches toward new samples. This works poorly in open areas (too slow) and dense obstacle fields (too risky). The new “AdjustableSteer” function starts with a base step but increases it incrementally in free space—speeding up coverage—then resets to a smaller value upon collision, allowing finer navigation through narrow passages. It’s adaptive, context-aware motion planning in real time.
Finally, during the optimization phase, instead of blindly rewiring every nearby node (as RRT* does), they implemented a heuristic filter called NodeOptimization. Before considering a new node for parent reassignment, the system estimates whether adding it could possibly improve the total path cost. Using a simple triangle inequality check—if the sum of distances from start to node and node to goal is less than the current best cost—the algorithm prunes unpromising candidates early. This reduces unnecessary computations and sharpens focus on truly beneficial rewires.
The cumulative effect is striking. In 2D simulations across three environments—empty, lightly obstructed, and a realistic citrus canopy map—the improved algorithm generated initial paths 78% faster than baseline Informed-RRT*, with 24% fewer nodes and 60% fewer iterations. After full optimization, those gains widened: 88% less time, 42% fewer nodes, and a 6% reduction in final path cost.
But simulation is just the warm-up. The real test came in 3D, inside the Robot Operating System (ROS) ecosystem, using MoveIt—an open-source motion planning framework widely adopted in research and industry. The team integrated their algorithm into OMPL (Open Motion Planning Library), the backend engine powering MoveIt, and tested it against three benchmarks: RRT-Connect, RRT, and standard Informed-RRT.
Each trial involved navigating the Aubo-i5 arm through a virtual orchard with four reconstructed citrus trees, each laden with spherical fruit models surrounded by cylindrical branch proxies. The robot began from a fixed home pose and had to reach designated fruits while maintaining a horizontal end-effector orientation—mimicking real-world gripper constraints.
Over 80 trials across four distinct tree configurations, the enhanced algorithm consistently outperformed all others. In the most challenging setup (Environment C, with dense inner-canopy obstacles), it achieved a 90% success rate—versus just 80% for Informed-RRT, 85% for RRT-Connect, and a dismal 65% for vanilla RRT. Planning time never exceeded 0.86 seconds, compared to over 1.6 seconds for the next-best performer. Path smoothness and node efficiency followed the same trend, indicating not just faster planning but higher-quality trajectories.
Why does this matter beyond academic metrics? Because in commercial agriculture, every second counts. A human picker might harvest 800–1,000 oranges per hour. To be economically viable, a robot must approach that throughput. If each pick takes 5 seconds—including perception, planning, motion, and release—then planning alone should ideally consume less than 1 second. The Chongqing team’s 0.81-second average puts them firmly in that window.
Moreover, reliability is non-negotiable. A 96% success rate means the robot rarely gets stuck or aborts—a crucial factor for unsupervised operation in large orchards. Combine that with reduced path length (lower energy use, less wear on joints) and smoother motion (less vibration, safer around fragile fruit), and you have a system that’s not just smart, but farm-ready.
Of course, challenges remain. The current implementation assumes static environments—no wind, no moving branches, no sudden occlusions. Future work will likely integrate real-time replanning and tighter coupling with vision systems for dynamic obstacle avoidance. Also, while the pre-pick point is fixed at 25 cm horizontally, optimal staging distance may vary by fruit type, tree architecture, or gripper design. Adaptive guide-point selection could be the next frontier.
Still, this research marks a significant pivot—from viewing path planning as a generic geometric problem to tailoring it to the biomechanical realities of agricultural tasks. By embedding domain knowledge (the need for staged approach) directly into the algorithm’s structure, the team achieved gains that pure computational tricks couldn’t deliver.
As global food demand rises and rural labor pools shrink, the pressure to automate harvesting intensifies. Solutions won’t come from brute-force hardware alone; they’ll emerge from clever software that understands the nuances of nature. This enhanced Informed-RRT* variant is a compelling example: a planner that doesn’t just find a way—but finds the right way, quickly, reliably, and gracefully.
For citrus growers watching the horizon, that’s not just progress. It’s promise.
Liu Dun, Wang Yi
College of Mechanical Engineering, Chongqing University of Technology, Chongqing 400054, China
Journal of Chongqing University of Technology (Natural Science), 2021, 35(11): 158–165
doi: 10.3969/j.issn.1674-8425(z).2021.11.020