The proposed invention is a TD3-based continuous control system for a two-wheeled self-balancing agricultural robot, designed to maintain dynamic stability, continuous forward motion, and straight-line trajectory tracking under varying soil traction conditions. The system addresses the effects of changing wheel–ground friction that can occur due to differences in agricultural soil conditions.

The robot is modelled using URDF and simulated in the PyBullet physics engine, which provides gravity, contact dynamics, and adjustable wheel–ground friction. Three friction coefficients, μ = 0.1, 1.0, and 1.5, are used to represent low-, medium-, and high-traction terrain conditions.

A custom Gym environment provides the TD3 controller with an 8-dimensional continuous observation state, consisting of pitch angle, pitch angular velocity, yaw angular velocity, cross-track error, forward velocity, lateral velocity, and left- and right-wheel velocities. The controller produces a 2-dimensional continuous action, directly controlling the left- and right-wheel velocities.

The control system employs an actor network and twin critic networks with a multi-objective reward function. The reward promotes upright balancing, forward movement, and accurate trajectory tracking while penalizing excessive tilt, yaw motion, and trajectory deviation. Each controller is trained for 2,000,000 simulation steps and subsequently evaluated for 2,000 steps under its corresponding friction condition.

Experimental results demonstrate that the proposed system maintains dynamic balance and continuous forward motion across all tested friction conditions while achieving reliable straight-line tracking. Training completion rates ranged from 77.16% to 86.51%, demonstrating the potential of TD3-based continuous control for agricultural robots operating under varying ground conditions.