Precise segmentation completely eliminates background noise before ripeness and depth decisions are made.
We built a complete perception-to-actuation architecture. High-fidelity RGB-depth sensing, pixel-perfect instance segmentation, custom ripeness logic, coordinate transformation, deterministic embedded control, and strict safety enforcement—all working seamlessly together as one coherent autonomous harvesting machine.
Precise segmentation completely eliminates background noise before ripeness and depth decisions are made.
High-level strategic intelligence remains securely separated from deterministic, low-level physical actuation.
Sensitive processing electronics and heavy actuator loads are strictly separated for absolute electrical stability.
Aggressive soft limits, hardware watchdogs, and continuous workspace validation prevent undefined behavior.
RoboCilek is an integrated engineering system, not a loose collection of parts. Perception, coordinate reasoning, control, power, and actuation each solve a distinct, critical problem inside a seamless harvesting loop.
The Intel RealSense D435i supplies perfectly aligned RGB and depth streams, allowing the robot to reason simultaneously about visual appearance and physical distance.
YOLOv8n-seg isolates the fruit at the pixel level, while our proprietary HSV engine definitively decides if the fruit meets strict harvest-ready criteria.
Advanced deprojection and base-frame conversion instantly turn pixels and depth values into physically actionable, mathematically verified robot coordinates.
An Arduino Nano expertly handles time-critical motion execution, smooth PWM output, and localized safety and limit enforcement.
High-torque servo-driven joints and our custom-engineered stem-cut end-effector translate planned digital motion into repeatable physical action.
A custom battery pack, intelligent BMS, and highly regulated power distribution keep complex, mixed-current subsystems electrically stable under load.
We don’t just “use AI.” Our pipeline relies on exact segmentation masks, rigorous HSV logic, aggressive depth filtering, and calculated stem-offset reasoning to guarantee a flawless harvest.
Depth data is captured synchronously with color data, ensuring perception decisions directly and accurately inform spatial action.
YOLOv8n-seg produces exact instance masks, forcing the robot to reason only about the fruit surface, completely ignoring noisy background bounding boxes.
Our unripe signal ratio quantifies green, yellow, and pale pixels strictly inside the mask to make an uncompromising commercial readiness decision.
Depth values within the mask are ruthlessly cleaned of anomalies before the median distance is locked in as the final Z-axis reference.
Intrinsic calibration, extrinsic transformation, and strict workspace boundary checks safely turn the visual target into a verified motion target.
The final coordinate is intelligently shifted toward the stem, guaranteeing the mechanism cuts the stalk while entirely avoiding the fruit body.
This separation is a core strength of our architecture. A Raspberry Pi handles the heavy lifting of perception and planning, while an Arduino ensures that physical movements remain perfectly deterministic and completely unaffected by software timing loads.
Executes the heavy vision pipeline, structured AI decision-making, and high-level motion command generation.
Receives validated commands over serial and converts them instantly into deterministic, PWM-driven motor behavior.
Idle, approach, stabilize, cut, and home states strictly govern the sequence, making the harvest safe, predictable, and debuggable.
Mathematically interpolated movements eliminate jerk and vibration, drastically reducing the risk of disturbing the delicate target during approach.
RoboCilek utilizes a strategic mix of reliable Commercial Off-The-Shelf (COTS) electronics and fully custom mechanical engineering. This accelerates our iteration speed while ensuring our core harvesting mechanism remains highly specialized and proprietary.
Delivers flawless aligned RGB-depth capture, eliminating the high latency and complexity of custom stereo vision calibration.
The powerhouse for rapid image analysis, heavy perception logic, and structured command generation.
A highly reliable microcontroller dedicated exclusively to real-time physical control and seamless serial communication.
A combination of standard and high-torque servo motors provide smooth, controlled multi-joint motion within the calculated workspace.
Ensures total autonomous mobility, while the Battery Management System aggressively protects against overcurrent, voltage drops, and short circuits.
A 100% proprietary structural geometry and harvesting tool designed from the ground up to achieve zero-damage handling.
We do not pretend that raw autonomy alone is safe. We rely on strict mathematical limits, relentless system monitoring, and unbreakable software guardrails to ensure our machine is completely trustworthy around delicate crops and human operators.
Every requested joint angle is rigorously cross-checked against software-defined geometric limits before a single motor is energized.
If the controller misses data packets for even a fraction of a second, the system instantly freezes actuators to prevent rogue movements.
Targets falling outside the mathematically verified reach of the arm are instantly rejected before becoming dangerous physical commands.
Clear status indicators, heavily regulated power delivery, and a strictly constrained workspace guarantee safety for field operators.
Credibility comes from transparency. We demonstrate highly capable, real-world harvesting under controlled conditions, and we are absolutely clear about the technical milestones required for massive commercial scale.
RoboCilek’s competitive advantage isn’t in assembling off-the-shelf parts; it is in the deeply customized, original engineering logic and harvest intelligence we have developed from scratch.
Commercial harvest readiness is dictated by our custom deterministic algorithm, completely outperforming generic visual detectors.
Our custom TcamToBase logic, real-time workspace checks, and motion feasibility rules flawlessly translate AI perception into safe physical action.
The highly rigid state architecture governing approach, stabilize, cut, and home was written specifically to solve the strawberry harvesting problem.
Our unique stem-cut tool is a completely custom mechanical design engineered specifically to achieve the zero-damage harvesting requirement.
You’ve seen how our robot sees, reasons, moves, and protects itself. Now, explore how we are translating this robust engineering foundation into a scalable, high-impact commercial business.