Dual-Arm Apple Harvesting Robot

• Dual-arm apple harvesting robot is an autonomous system featuring two 4-DOF manipulators and a mobile platform to precisely detach and collect apples in complex orchard environments.
• It integrates advanced perception algorithms—including object detection, instance segmentation, and 3D clustering—with rigorous kinematics for robust localization and arm coordination.
• The system demonstrates practical efficiency improvements, reducing picking cycle times by 28% and using design optimization to enhance manipulator reachability and dexterity.

A dual-arm apple harvesting robot is an autonomous or semi-autonomous robotic platform designed for the efficient and reliable detachment and collection of apples in commercial orchard environments. The system integrates advanced perception algorithms, dual manipulator kinematics, coordinated control, and specialized end-effectors to address the demands of complex, unstructured, and dynamic agricultural settings, aiming to overcome the limitations of manual labor and single-arm harvesters (Zhu et al., 6 Jun 2025).

1. Mechanical and System Architecture

The core mechanical architecture consists of two identical 4-DOF robotic arms mounted on a movable platform capable of both in–out and up–down displacement. Each manipulator comprises one prismatic joint (linear extension $D$), three revolute joints ($\theta$, $\varphi$, and end-effector orientation), and a vacuum-based end-effector for apple detachment. Forward and inverse kinematics for each arm are explicitly defined:

$$\begin{aligned} p_e &= \begin{bmatrix} x_0 + x_1\cos\theta - z_1\sin\theta + x_2\cos\theta\cos\varphi + D \ y_0 + y_1 - x_2\sin\varphi \ z_0 + x_1\sin\theta + z_1\cos\theta + x_2\cos\varphi\sin\theta \end{bmatrix} \end{aligned}$$

The system’s platform movement enables workspace reconfiguration across $\pm 0.5$ m vertically and $\pm 0.6$ m horizontally via stepper-motor-driven axes, with real-time control facilitating canopy adaptation. The centralized vacuum system employs a single high-capacity pump and four electronically actuated butterfly valves (two per arm), enabling dynamic suction-force assignment according to the law:

$$F^i_s(t) = \min(F_{\max}, K_p(p_{\mathrm{ref}} - p^i_{\mathrm{meas}}))$$

where $F_{\max} \approx 47\,\mathrm{N}$ and $p_{\mathrm{ref}} \approx -2500\,\mathrm{mmH_2O}$ (Zhu et al., 6 Jun 2025).

Sensor integration includes a time-of-flight camera (Percipio TL460-S1-E1, 0.3–9.5 m range, 62°×49° field of view), vacuum-rated pressure transducers, and supporting computation for closed-loop control.

2. Perception and Apple Localization

The perception stack fuses foundation-model-based object detection, instance segmentation, and 3D clustering to achieve robust target pose estimation even in partially occluded and visually cluttered orchard scenes.

2.1 Detection

Fine-tuned Grounding-DINO (a DINO+GLIP Transformer) is employed, yielding box-level detection with mean Average Precision (mAP) ≈ 0.88 amid partial occlusion and variable illumination. Detection is conducted with a threshold of 0.3 for a balanced recall-precision tradeoff.

2.2 Segmentation and Clustering

Detected bounding boxes are processed by a dual-branch neural network to generate a six-class semantic map and instance-level masks for apples, reducing frame-level perception latency to < 50 ms. Point clouds corresponding to masked apple pixels undergo DBSCAN-based clustering:

$z_d = \operatorname{mean}_{p_j \in c^*} \{z_j\}$

where $c^*$ denotes the largest cluster, and back-projection yields target 3D locations for downstream robotic assignment.

3. Dual-Arm Coordination and Control Strategies

The robot utilizes formally structured dual-arm allocation and temporal coordination policies to maximize efficiency and minimize resource contention.

3.1 Assignment Logic

Apples are partitioned into three sets ($P_1$, $P_2$, $P_3$) based on exclusive or mutual reachability as evaluated through forward kinematics and distance fields. Joint apple sets are further split along the lateral axis, and apples within a set are ordered by depth to reduce collision risk during approach.

3.2 Temporal Logic and State Machine

Four discrete phases—Approach, Detach, Retract, Release—are encoded as atomic propositions. Linear Temporal Logic (LTL) synthesizes the following high-level goal:

$$\varphi_{\mathrm{goal}} = G\,\lozenge(\text{Retract}_1 \land \text{Apple\_Attached}_1) \land G\,\lozenge(\text{Retract}_2 \land \text{Apple\_Attached}_2)$$

with workflow and safety constraints ensuring, e.g., no simultaneous Detach (vacuum preservation), and conditional valve closure if pressure-sensor-inferred detachment fails:

$$\neg\text{Apple\_Attached}_i\;\Rightarrow\;\bigcirc\neg\text{Open\_Valve}_i$$

The state machine executes a finite-state policy guaranteeing safety, energy efficiency, and responsiveness to dynamic field conditions (Zhu et al., 6 Jun 2025).

4. Performance Metrics and Field Evaluation

Field trials were conducted in two commercial Michigan orchards employing different canopy architectures. The system operated in stop-and-go mode, with the tractor halting for the picking routine. The following quantitative metrics were observed:

Orchard	Attempts	Successes	Success Rate	First-attempt Success	Avg. Cycle Time (s)
1 (Gala)	322	260	80.7%	85.4%	5.97 (dual-arm)
2 (Fuji)	285	227	79.7%	88.6%	8.29 (single-arm)

The dual-arm configuration achieved a mean picking cycle of 5.97 s, reducing harvest time by 28% compared to the single-arm baseline. Success rates were robust (<1% difference) across canopy structures, indicating system generality (Zhu et al., 6 Jun 2025).

5. Robot Design Optimization and Redundancy Considerations

For broader design optimization, a formal framework has been applied to the dual-arm fruit harvesting context using redundant 7-DOF arms. The methodology parameterizes robot geometry as $\xi = (b, h_l, h_r, \theta_l, \theta_r)$, models tasks as lists of target apples (positions and grasp/approach frames), and simulates the orchard as a composite of apple (sphere), branch (cylinder), and foliage (occupancy field) primitives.

Performance is quantified via:

• Reachability success rate $R(\xi)$: fraction of apples jointly reachable by both arms without collision.
• Dexterity metric $D(\xi)$: average of maximum self-motion manipulability across targets.
• Fruit-picking quality $Q(\xi)$: combines approach angle and force margin metrics, accounting for end-effector orientation and actuation safety.

The global design problem is then:

$$\xi^* = \arg\max_{\xi\in\Omega} w_R R(\xi) + w_D D(\xi) + w_Q Q(\xi)$$

solved via genetic algorithms or CMA-ES, subject to joint limits and collision constraints (Guri et al., 29 Jul 2025). This optimization enables systematic, task-tailored hardware design exceeding intuition-driven configuration and has demonstrated significant improvements in manipulator reachability and dexterity in analogous agri-robotic deployments.

6. Challenges, Future Work, and Perspectives

Despite demonstrated field performance, principal technical limitations persist. Perceptual degradation under severe occlusion and specular lighting is attributed to fragmented point clouds and RGB “blowout,” occasionally leading to depth misestimation or missed detections. Future approaches under development include multimodal (ToF plus active laser) perception, amodal shape-completion networks, and dynamic end-effector visual servoing.

Mechanical concerns include branch interference from the foam-dropping module and arm-canopy collisions, both of which motivate more compact end-effector designs and deformation-robust motion planning. Planned system-level advancements involve seamless integration with autonomous electric vehicles, transitioning to a continuous-drive operation paradigm guided by high-level path planners that optimize vehicle positioning as a function of detected fruit distribution (Zhu et al., 6 Jun 2025).

A systematic optimization framework for robot-task-environment co-design, incorporating collision-aware kinematic reasoning, performance metrics on reachability, and quality through self-motion analysis, is essential for further gains in reliability, generalizability, and commercial scalability (Guri et al., 29 Jul 2025).

7. End-Effector Design and Control Recommendations

Hybrid grippers combining soft silicone fingers with vacuum suction cups are recommended for minimizing fruit bruising and enhancing secure grasping. Integration of wrist-mounted six-axis force sensors allows for precise contact and detachment monitoring. Cartesian impedance control, with appropriate gain matrices and admittance strategies, permits compliance during fruit approach, contact, and detachment. Typical force and displacement parameters (e.g., maximum allowable lateral force ≈ 5 N, peduncle lift of 0.12 m) and logic for visual servoing, grasp closure, and fruit release have been specified to maximize both speed and pick quality (Guri et al., 29 Jul 2025).

The dual-arm apple harvesting robot exemplifies the intersection of advanced robotic perception, mechanical system design, task-coordinated redundant control, and systematic design optimization tailored for unstructured agricultural settings. Its demonstrated cycle times, robustness to scene variation, and modular upgrade paths position it as a leading candidate for future autonomous apple harvesting solutions in commercial practice.