Recent advances in legged robotics have highlighted deep reinforcement learning (DRL)-based controllers for their robust adaptability to diverse, unstructured environments. While position-based DRL controllers achieve high tracking accuracy, they offer limited disturbance rejection, which degrades walking stability; torque-based DRL controllers can mitigate this issue but typically require extensive time and trial-and-error to converge. To address these challenges, we propose a Real-Time Motion Generator (RTMG). At each time step, RTMG kinematically synthesizes end-effector trajectories from target translational and angular velocities (yaw rate) and step length, then maps them to joint angles via inverse kinematics to produce imitation data. The RL agent uses this imitation data as a torque bias, which is gradually annealed during training to enable fully autonomous behavior. We further combine the RTMG-generated imitation data with a decaying action priors scheme to ensure both initial stability and motion diversity. The proposed training pipeline, implemented in NVIDIA Isaac Gym with Proximal Policy Optimization (PPO), reliably converges to the target gait pattern. The trained controller is Tensor RT-optimized and runs at 50 Hz on a Jetson Nano; relative to a position-based baseline, torso oscillation is reduced by 24.88% in simulation and 21.24% on hardware, demonstrating the effectiveness of the approach.
A wheeled–legged robot has the advantage of stable and agile movement on flat ground and an excellent ability to overcome obstacles. However, when faced with a narrow footprint, there is a limit to its ability to move. We developed the control moment gyroscope (CMG) unicycle–legged robot to solve this problem. A scissored pair of CMGs was applied to control the roll balance, and the pitch balance was modeled as a double-inverted pendulum. We performed Linear Quadratic Regulator (LQR) control and model predictive control (MPC) in a system in which the control systems in the roll and pitch directions were separated. We also devised a method for controlling the rotation of the robot in the yaw direction using torque generated by the CMG, and the performance of these controllers was verified in the Gazebo simulator. In addition, forward driving control was performed to verify mobility, which is the main advantage of the wheeled–legged robot; it was confirmed that this control enabled the robot to pass through a narrow space of 0.15 m. Before implementing the verified controllers in the real world, we built a CMG test platform and confirmed that balancing control was maintained within .
The Single Inverted Pendulum (SIP) model is widely employed in various robotics applications for balance control due to its structural simplicity and ease of implementation. Self-balancing two-wheeled transport robots offer the advantage of maintaining stability while enabling dynamic acceleration and deceleration. However, their performance is limited when carrying heavy loads, and maintaining stable balance during the loading process remains challenging. To address these limitations, this study proposes a control strategy for a pair of self-balancing two-wheeled transport robots that operate cooperatively through a separation–combination mechanism. Simulations were conducted to compare the performance of the Linear Quadratic Regulator (LQR) and Model Predictive Control (MPC) under a self-balancing condition with an external force of 100 N applied. Additionally, the proposed separation–combination method was evaluated through a simulation involving the transfer of a 20 kg delivery module. The results showed that both methods could stably transport the load, while the MPC-based approach reduced the wheel torque by approximately 50% compared to the LQR method. These findings indicate that a two-wheeled robot employing MPC in conjunction with the separation–combination method is well-suited for transporting heavy loads.