Bark Watch

Our proposed system combines advanced reinforcement learning, robust mechanical design, and a modular control architecture to deliver reliable field performance. The quadruped uses rotary BLDC actuators selected for their torque density and smooth motion, enabling a lightweight, deer-leg-inspired morphology optimized for mud, roots, and uneven ground. A Jetson Orin Nano performs high-level tasks—including state estimation, trajectory generation, and execution of a trained policy—while a Raspberry Pi with distributed CAN FD networks handles low-level actuator control at 1 kHz. This separation ensures real-time responsiveness and scalability.
In simulation, our reinforcement learning framework trained 6,000 virtual quadrupeds in parallel using domain randomization to improve Sim2Real transfer. The agents demonstrated strong command tracking, stable gait patterns, and emergent behaviors such as balance recovery, diagonal mirroring, and adaptive stance control. Although the initial model will require retraining using the finalized robot geometry and sensing strategy, the developed pipeline significantly accelerates future iterations.
The design meets key user-driven specifications including >0.6 m/s traversal speed, one-hour swappable battery operation, operation on 20% grade terrain, and the ability to move through 0.5 m undergrowth without obstructing sensors. With a target weight under 30 kg and a projected operating cost of < $1,500 per hectare per year, the system aims to dramatically reduce the cost and improve the accuracy of reforestation monitoring.