Fruit Picking Robot

Overview

The Fruit Picking Robot is an autonomous mobile manipulator designed for selective harvesting of tree fruits in commercial orchards. Using dual 6-axis arms with soft pneumatic grippers and integrated stereo vision, it identifies ripe fruit, navigates to the target, grasps without bruising, and delivers the harvest to an onboard conveyor system. The system uses lightweight materials and force-feedback sensors to handle delicate produce safely while maintaining high throughput.

The platform combines three core technologies: a wheeled mobile base for orchard navigation, dual collaborative robot arms for dexterous manipulation, and edge AI for ripeness detection. Power comes from a 480 Wh LiFePO4 battery pack with multi-voltage regulation. The system is designed for outdoor operation in variable terrain and can work alongside human crews or autonomously in designated harvest zones.

How it works

The robot operates in a perception-plan-execute loop:

Vision and Ripeness Detection

Dual RGB cameras mounted on the head provide stereo depth estimation, while an RGB-D sensor captures precise 3D geometry of branches and fruit. The onboard NVIDIA Jetson Nano runs a CNN-based ripeness classifier trained on spectral analysis of fruit color and shape. The AI processor achieves 95%+ accuracy for target species and ripeness state at 15 fps. Confidence thresholds above 0.85 trigger arm movement; below that, the robot navigates to the next candidate.

Navigation and Arm Positioning

A STM32H7 microcontroller orchestrates the mobile base, feeding wheel encoder feedback into a simple wheel odometry tracker. LIDAR or sonar-based obstacle avoidance keeps the base centered within 30 cm of the target fruit. Once positioned, the left and right arms move in parallel: the left arm approaches from one side while the right stabilizes the branch. Each arm has six degrees of freedom (three for reach, three for orientation), driven by servo motors from 10 Nm to 90 Nm depending on joint.

Soft Gripper Control

Each gripper has three silicone fingers actuated by pneumatic pistons. Pressurized air at 4-6 bar closes the fingers around fruit; a pressure transducer ensures grip force stays below 5 N to avoid bruising. A proportional solenoid valve modulates air flow for dexterous grasping and release. The gripper can handle fruit diameters from 4 cm to 12 cm.

Fruit Conveyance

Once picked, fruit slides down a short chute onto an articulated collection tray. A brushless conveyor belt moves the fruit toward a central bin. When the tray is full (sensed by load cell), a servo motor tilts the tray to dump fruit into a larger container. The system can harvest 60-80 pieces per hour in dense-canopy scenarios.

Mechanical Design

The mobile base uses four pneumatic 6-inch wheels with independent brushless motor drives, allowing skid-steering for tight turns between trees. The chassis is welded aluminum with a 15 cm ground clearance for rough orchard soil.

The dual arms are serial-chain structures built from carbon fiber tubes and aluminum joints. Each joint uses a servo motor with integral feedback, allowing closed-loop position control at 50 Hz. The arm links are sized for a 1200 mm reach, sufficient to grasp fruit on the inner canopy while the base remains 60 cm from the trunk.

The gripper actuator is a compact proportional valve integrated into the wrist. Soft fingers use strain relief design to prevent sharp creases that could bruise fruit. Testing shows minimal damage even at impact speeds of 0.2 m/s.

Electrical and Control Architecture

The power system is a 48V 10 Ah LiFePO4 pack with integrated BMS and active cell balancing. The BMS monitors cell voltage, temperature, and current, shutting down if any cell exceeds 4.2V or temperature rises above 55°C. Two DC-DC converters step down to 12V for arm servos and 5V for sensors and wireless modules.

The main controller is a STM32H7-based board running FreeRTOS. It samples arm encoders, gripper pressure transducers, and wheel odometry at 50 Hz. Commands are received over WiFi 6 from a cloud planner or local field operator, with Bluetooth fallback if WiFi is unavailable.

The vision pipeline runs on NVIDIA Jetson Nano at 15 fps. Raw RGB frames are processed through a ResNet-50 backbone trained on 50,000 labeled fruit images. The model outputs bounding boxes and ripeness scores; only boxes with confidence > 0.85 and ripeness in a specified window (e.g., color saturation > 70%) trigger arm movement.

Deployment Considerations

Orchard operators deploy the robot in rows, typically 3-5 units per acre. Each unit communicates with a central fleet controller via WiFi mesh, allowing operators to monitor battery status, harvest rate, and equipment faults in real time. The system is designed to work safely alongside humans; the arms have force-limiting at 150 N to trigger stop if a person is struck.

Maintenance intervals are 40 operating hours for pneumatic filter changes, 200 hours for gripper finger inspection, and 500 hours for servo bearing inspection. Battery replacement is recommended at 1000 full cycles (roughly 2 years of 5-day-per-week operation).