Multi-Arm Harvesting Robot Breaks Efficiency Barrier in Dwarf Orchards
In a significant leap forward for agricultural robotics, researchers from Beijing Academy of Agriculture and Forestry Sciences have developed a novel task-planning framework that dramatically enhances the efficiency of multi-arm harvesting robots in high-density dwarf orchards. The breakthrough, published in Transactions of the Chinese Society of Agricultural Engineering, introduces a new computational model—Asynchronous Overlapped Multiple Traveling Salesman Problem (AOMTSP)—to address one of the most persistent challenges in robotic agriculture: coordinating multiple robotic arms in shared workspaces without collisions, while minimizing total operation time.
The research, led by Dr. Li Tao and Dr. Qiu Quan from the Beijing Research Center of Intelligent Equipment for Agriculture, in collaboration with Dr. Zhao Chunjiang from the Beijing Research Center for Information Technology in Agriculture and Dr. Xie Feng from Jiangsu University, tackles the inefficiencies that have long plagued robotic fruit harvesting systems. While single-arm robots have struggled to match human picking speeds, multi-arm systems offer a promising path to scalability. However, their performance has been hindered by complex coordination issues, especially when arms operate in overlapping zones.
Traditional task-planning methods have relied on either strict spatial partitioning—where each arm is assigned a fixed zone to avoid collisions—or simple random assignment. The former often leads to “dead zones” where fruit is unreachable, while the latter increases the risk of mechanical interference and inefficient travel paths. The team’s new approach redefines the problem by integrating the overlapping access domain into the planning process, rather than treating it as a constraint to be avoided.
At the core of the innovation is the AOMTSP model, a sophisticated adaptation of the classic Traveling Salesman Problem (TSP), which seeks the shortest possible route for visiting a set of locations. In multi-arm systems, the Multiple TSP (MTSP) has been used to assign routes to several agents. However, MTSP assumes non-overlapping territories, making it unsuitable for robots with shared workspaces. The AOMTSP model overcomes this limitation by allowing designated overlap zones where multiple arms can operate, but only one at a time, thus preventing collisions while ensuring full coverage of the orchard.
To solve the AOMTSP, the researchers employed a modified genetic algorithm (GA), a powerful optimization technique inspired by natural selection. The GA was tailored to handle the dual challenges of route optimization and conflict avoidance. A key feature of the algorithm is its dual-chromosome encoding scheme. One chromosome encodes the sequence of fruit targets, while the other assigns each target to a specific robotic arm. This structure enables the algorithm to simultaneously optimize both the path and the allocation of tasks, ensuring that no two arms are scheduled to enter the same overlap zone simultaneously.
The asynchronous rule embedded in the algorithm introduces a virtual “queuing time” when two or more arms attempt to access the same shared zone. This queuing time is treated as a penalty in the fitness function, guiding the evolutionary process toward solutions that minimize both travel distance and waiting periods. The result is a globally optimized task plan that balances speed and safety.
The team tested their AOMTSP-GA framework on a custom-built four-arm Cartesian robot designed for dwarf apple orchards. The robot features shared X-axis rails and adjustable Y- and Z-axis rails to create overlapping workspaces, eliminating dead zones. In simulation trials, the system was tasked with harvesting 43 and 90 fruit targets, representing average and high-density scenarios in real orchards.
The results were striking. For 43 fruits, the AOMTSP-GA algorithm converged within 500 iterations, reducing the total traversal time by 40.97% compared to a random traversal method. When scaled to 90 fruits, the improvement reached 54.98%, with convergence achieved in 2,000 iterations. The algorithm also outperformed single-arm systems by a factor of 4.28, demonstrating the clear advantage of coordinated multi-arm operation.
But the real test came in physical experiments using a dual-arm version of the robot. In three distinct fruit distribution scenarios—sparse, moderately clustered, and highly concentrated in the overlap zone—the AOMTSP-GA method consistently reduced traversal time. Compared to sequential planning and random traversal, the new method shortened operation time by 10.69% and 27.18% in the sparse case, 20.45% and 23.33% in the moderate case, and 12.94% and 21.69% in the dense case.
The researchers noted a critical insight: in sparse distributions, path length is the dominant factor in efficiency, while in dense clusters, conflict avoidance becomes paramount. The AOMTSP-GA framework excels in both regimes by dynamically balancing route optimization and access coordination. In the densest scenario, where overlap zones contained 20 of 28 fruits, competing methods suffered from frequent collisions and long waiting times, while the AOMTSP-GA solution maintained smooth, conflict-free operation.
The implications of this work extend far beyond apple orchards. High-density dwarf cultivation is rapidly gaining traction worldwide due to its higher yields, easier management, and compatibility with mechanization. From cherries to citrus, many fruit crops are being restructured into “fruit wall” architectures that resemble the conditions tested in this study. As such, the AOMTSP-GA framework offers a generalizable solution for any multi-arm harvesting system operating in constrained, shared environments.
Moreover, the integration of the algorithm into a real-time ROS (Robot Operating System) framework demonstrates its practical viability. The system uses NVIDIA Jetson TX2 for edge computing, processing depth data from Intel RealSense D455 cameras to detect and localize fruit in 3D space. Once targets are identified, the AOMTSP-GA planner computes optimal sequences for each arm, which are then executed via MoveIt! motion planning. This end-to-end pipeline—from perception to action—runs autonomously, enabling continuous harvesting without human intervention.
The success of this project underscores a broader shift in agricultural robotics: from isolated hardware innovations to integrated, intelligence-driven systems. While early efforts focused on building robust grippers or accurate vision systems, the next frontier lies in coordination and decision-making. As farms scale up and labor shortages intensify, the ability to orchestrate multiple robots in complex environments will be a decisive factor in the adoption of automation.
Dr. Qiu Quan, the corresponding author, emphasized that the team’s approach is not just about speed, but about reliability and scalability. “In real-world orchards, every second counts, but so does safety,” he said. “Our method ensures that the robot can operate continuously, without collisions or missed fruit, even in the most challenging conditions.”
The research also highlights the importance of co-design between hardware and software. The robot’s mechanical structure—specifically the shared X-axis and rotatable Y- and Z-axes—was engineered to enable overlap, which in turn enabled the AOMTSP model to function. This synergy between physical design and algorithmic innovation is a hallmark of effective robotics research.
Looking ahead, the team plans to expand the framework to accommodate dynamic environments, where fruit positions may change due to wind or branch movement, and to integrate real-time replanning capabilities. They are also exploring the use of reinforcement learning to further refine the task allocation process.
The commercial potential is substantial. Companies like Agrobot and FFRobotics have already deployed multi-arm harvesters, but their systems often rely on simpler, less adaptive planning methods. The AOMTSP-GA framework could provide a competitive edge by enabling faster, more reliable harvesting with fewer mechanical failures.
For farmers, the benefits are clear: reduced labor costs, higher yields, and the ability to harvest at optimal times without being constrained by workforce availability. In regions facing acute labor shortages, such as California’s Central Valley or Spain’s Murcia region, robotic harvesters could become indispensable.
The study also contributes to the growing body of work on multi-agent systems in unstructured environments. While much of the research in multi-robot coordination has focused on structured settings like warehouses or factories, orchards present unique challenges: uneven terrain, variable lighting, occlusions from leaves, and irregular fruit distribution. Solving these problems requires not just better sensors or stronger actuators, but smarter planning algorithms that can reason about space, time, and cooperation.
The AOMTSP model represents a step toward more autonomous agricultural robots—machines that don’t just follow pre-programmed paths, but actively plan and adapt to their surroundings. As the global population grows and food demand increases, such technologies will be essential for sustainable food production.
In an era where artificial intelligence is transforming industries from healthcare to finance, agriculture has often been overlooked. But this research shows that AI can be just as powerful in the field as it is in the cloud. By combining advanced robotics, computer vision, and evolutionary algorithms, the team has created a system that not only picks fruit faster but does so with a level of coordination and intelligence that was previously unattainable.
The work also reflects a broader trend in Chinese agricultural research: a shift from incremental improvements to bold, systems-level innovations. With strong government support for smart agriculture and a growing pool of engineering talent, China is emerging as a leader in agri-tech, particularly in robotics and AI applications.
As the world grapples with climate change, resource scarcity, and food security, the need for efficient, scalable farming solutions has never been greater. The AOMTSP-GA framework is not just a technical achievement—it is a blueprint for the future of farming, where intelligent machines work in harmony to feed a growing planet.
The research was supported by the National Natural Science Foundation of China, the China Postdoctoral Science Foundation, and the Beijing Academy of Agriculture and Forestry Sciences Postdoctoral Research Fund. The full paper, including detailed methodology and experimental results, is available in Transactions of the Chinese Society of Agricultural Engineering.
Li Tao, Qiu Quan, Zhao Chunjiang, Xie Feng. Multi-Arm Harvesting Robot Task Planning for High-Density Dwarf Orchards. Transactions of the Chinese Society of Agricultural Engineering. doi:10.11975/j.issn.1002-6819.2021.2.001