Development of software (ROS 2 / Nav2) for a 4x4 autonomous robot: Computer vision, asymmetric navigation

Concept and kinematics of the device: A universal all-wheel drive 4x4 wheeled chassis is being developed (independent motor-reducers for each wheel with current sensors, passive balancing suspension). The mower blade is extended to the front cantilever deck — in front of the front wheel line with an asymmetric offset to the right. This geometry allows mowing grass in front of the wheels (avoiding trampling) and getting under the branches of bushes (blueberries, raspberries) near the trunks with a low console. Critical technical requirements for the software: Control interface WITHOUT mobile applications: All interaction with the operator occurs exclusively through a physical industrial information display with high contrast (OLED / LED / E-ink or LED matrix) and large mechanical buttons ([START], [YES], [NO], [FULL MOWING]) directly on the robot's body (other options may exist). The display must ensure clear reading of information in direct sunlight. No smartphones, internet, or app downloads in the field are provided. Camera reconnaissance mode: When entering a new area, the robot makes a test pass idling (blade off). The main industrial AI camera based on a microcomputer with an AI accelerator recognizes the crop (raspberries, blueberries, strawberries, etc.) and displays the message on the robot's display: "PROFILE FOUND: RASPBERRY. START?". After pressing the [YES] button, the blade is activated, and work begins according to the algorithm of smooth weaving between bushes. When a "bald spot" is detected, the blade does not retract, and the area is mowed completely. Locking and "Full Mowing" mode: If the AI does not recognize the crop due to the chaos of weeds (confidence <85%), the software displays the message "ERROR: UNKNOWN ROW" and blocks the operation. The operator can forcibly press the [FULL MOWING] button. The AI disables plant recognition, and the machine operates on a linear trajectory (clearing shrubs), relying only on current sensors in the wheel motors (stopping upon physical resistance against a stone). Universal AI recognition of obstacles and safety: The main front camera is the sole safety control organ. The neural network must clearly differentiate soft biomass (grass, weeds) from any artificial hard obstacles (metal pipes, concrete posts, wooden stakes) and living beings (cats, dogs, humans). Upon detecting any type of post/obstacle, the AI preemptively retracts the blade actuator by 15–20 cm to circumvent the obstacle in an arc. Upon detecting a human, cat, or dog in the field of view, the software instantly de-energizes the blade motor and locks the wheels, stopping the machine. The blade operates strictly while moving forward; during turns, it is completely de-energized. Defining field boundaries using markers (WITHOUT GPS): To mark invisible field boundaries, the operator sets one or two bright reflective markers (stakes or tape). The camera recognizes this color/reflection, builds a virtual fence line, and when approaching it, the robot turns 180 degrees. Following the "driven track" (vSLAM): The navigation algorithm for subsequent passes orients itself based on the visual texture of the edge of the already mowed strip of grass. When the 4x4 wheels are slipping in mud, the software detects slippage from the camera image, ignores false wheel data, and maintains a precise course to the end of the row. Energy-independent hot auto-resume (Resume Mission): The map and work profile built in the morning are automatically saved on the board. In the event of sudden power loss (battery replacement), the software does not crash due to the Read-Only file system. Upon turning on, the display shows the message: "CONTINUE CURRENT TASK?". After pressing the [YES] button, the robot resumes mowing strictly from the stop point. Smart tool recognition and protection: Loading the required AI profile ("Garden" or "Border/Snow") occurs automatically when closing the corresponding contacts (GPIO) in the power connector of the attachment cable. While the connector is not inserted until clicked — the power relay completely de-energizes the wheel motors (protection against accidental activation). Staged development and control (STRICTLY THROUGH SAFE): The development cost is discussed with candidates based on their proposals ("Price negotiable"). Payment for each stage is released to the performer only after successful demonstration of the result:

STAGE 1: Specification of hardware and simulation in software. The performer provides an exact list of industrial camera models and computing boards within our budget, justifying their choice. They configure the logic in the simulator (Gazebo or similar). Acceptance criterion: Stress test live via Zoom/Discord. We provide the performer with our personal video from the phone (where stakes, sticks, posts, or animals lie in the grass), and their AI on the computer must stably track objects in real-time and issue logical commands to the wheels/blade actuator without software hangs. STAGE 2: Hardware flashing on the table. Transferring software to the physical board and connecting the camera. Testing the board's response on the information display live. STAGE 3: Field tests, code transfer, and warranty. Calibrating the AI's reaction speed directly on our 4x4 wheeled chassis. Transferring clean source code (Open Source) with text comments for independent audit. Warranty support — 6 months.

In your project response, please specify:

What specific hardware architecture (platform, processor, camera module) do you propose to use for these requirements and why? Do you have experience working with ROS 2 / Nav2, Behavior Trees, and configuring Linux in Read-Only? Your total cost and timelines for each of the three stages separately.