Design Overview of KLEIYN

KLEIYN is a quadruped robot with a total of 13 degrees of freedom (DOF), consisting of 3-DOF per leg and 1-DOF in the torso. The robot weighs 18 kg, with a body length of 760 mm and a standing height of 400 mm. The torso consists of two identical components: the front-body-link and the back-body-link. Each body link has a box-shaped structure made of four aluminum sheet-metal plates, housing internal components such as the onboard PC and battery. These two body links are connected by the waist-joint, which allows pitch-axis bending. The leg design is based on MEVIUS, an open-source metal quadruped robot. Each leg consists of three links: the scapula-link, thigh-link, and calf-link, which are connected by three joints: the collar-joint, hip-joint, and knee-joint. All leg joints are actuated by motors with a 1:10 reduction ratio and a maximum torque of 25 Nm.

Waist Design

KLEIYN features a 1-DOF rotational joint along the pitch axis, allowing the torso to bend. The design of the waist joint is shown in Fig. 3. The frame is made of machined aluminum parts, and its strength was verified through pre-simulation stress analysis. To enhance rigidity, a double-supported structure is adopted, where a bearing on the opposite side of the motor output shaft supports the load. For easier sim-to-real transfer in reinforcement learning, the motor output is transmitted to the rotation axis via a low-reduction (1:9) gear in a quasi-direct-drive configuration.

System Architecture of Chimney Climbing

During training, two networks are trained by PPO: an Actor that outputs actions and a Critic that estimates value functions. When operating the real robot, only the Actor network is used. The Actor outputs scaled target joint angles at a frequency of 50 Hz. These target joint angles are transmitted via CAN communication to each joint motor driver, where they serve as reference values for the internal position control of each motor driver. KLEIYN’s sensors include encoders in each motor, a 3D LiDAR (Livox MID-360) mounted on the front-body-link, and an IMU inside the LiDAR. From the encoders, the joint angles (13 dims) and joint angular velocities (13 dims) are obtained. From the IMU, angular ve locity (3 dims) and estimated orientation (4 dims) are acquired. Additionally, a target vertical velocity vref (1 dim) is received from a Bluetooth controller. Although, the point cloud data obtained from the 3D LiDAR is used for SLAM with Fast-LIO, the localization results are only used for recording of z-position changes.

Reinforcement Learning for Chimney Climbing

Wall-climbing learning was conducted in a simulation environment using Isaac Gym. To enable the robot to climb walls of various widths, we prepared walls ranging from 900 mm to 1100 mm in width and used them for training. Additionally, as a domain randomization, we randomized the coefficient of friction of the feet within the range of 0.7 to 0.95 and introduced variations in the mass and moments of inertia of each link. To enhance robustness against external disturbances, we periodically applied random velocity perturbations of 0 to 1 m/s to the front-body-link and continuously applied random external forces and torques. The initial position of the robot was set to a standing posture on the ground (pz = 0.4 m), and the wall height in the simulator was set to 4 m. If the robot successfully climbed the wall, if z-position exceeds 3 m height, it was rewarded and reset to the initial position. We trained policy for 20,000 iterations in simulation and use it for control of real robot without any fine tuning.

Contact-Guided Curriculum Learning (CGCL) for chimney climbing

The key feature of chimney climbing learning is Contact-Guided Curriculum Learning (CGCL), where the junction between the wall and the floor gradually transitions from a curved surface to a vertical surface. Since the initial state starts from a standing position on the ground, KLEIYN have to learn transitioning motion from a standing posture to a bracing posture. However it requires the agent to experience rewards for successful bracing, which can take a considerable amount of time due to the random exploration nature of the learning algorithm. To accelerate learning, we designed an environment where, in the early stages of training, the terrain naturally induces the agent to experience bracing by randomly moving its feet. For this purpose, we utilized a curved junction represented by an elliptical arc with a vertical radius of 1.0 m and a horizontal radius of r. By adjusting r, the shape can be made smoother or more vertical. Initially, we set r = 0.3 m and gradually reduced it to r → 0.0 m as training progressed, enabling a smooth transition to fully vertical climbing. Additionally, to ensure adaptability to non-flat wall surfaces, the roughness of the wall was increased as learning progressed.

Experiments

Chimney Climbing in Simulation

Based on the learned policy, wall climbing motions were performed in the simulator for walls of width 750 mm, 900 mm, and 1100 mm at vref is 0.5 m/s. The climbing motion is realized by alternately executing two phases: the stance phase, where the body is pushed up, and the swing phase, where the legs are lifted. A significant upward velocity is momentarily generated in the vertical direction during the transition from the stance phase to the swing phase, indicating that the robot utilizes recoil when lifting its legs. Although the climbing speed decreases, the robot successfully climbs even the 750 mm-wide wall, which was not included in the training environment. This demonstrates the policy’s generalization capability to unseen wall widths. Furthermore, KLEIYN has a total length of 760 mm, making it impossible to climb a 750 mm-wide wall without utilizing its waist joint. This indicates that the robot has acquired the ability to climb narrower walls by leveraging its waist joint.

Chimney Climbing in Real Robot

To verify that the learned wall-climbing behavior could be applied to a real robot, we conducted physical experiments. We set up a pair of plywood walls facing each other and tested climbing on walls of three different widths: 800 mm, 900 mm, and 1000 mm. The robot was commanded to climb with a target velocity of vref 0.5 m/s. The climbing motion was achieved through alternating stance and swing phases, similar to the simulation experiments. Furthermore, recoil-based movement, observed in the simulation, was also reproduced in the real-world experiments. Additionally, slipping of the legs was frequently observed during the stance phase. In such cases, the robot quickly repositioned the slipped leg and reestablished bracing. This suggests that the learned policy is robust against disturbances, such as slipping, through reinforcement learning.

Bibtex

@inproceedings{yoneda2025kleiyn,
  title={{KLEIYN : A Quadruped Robot with an Active Waist for Both Locomotion and Wall Climbing (in press)}},
  author={Keita Yoneda and Kento Kawaharazuka and Temma Suzuki and Takahiro Hattori and Kei Okada},
  booktitle={2025 IEEE-RAS International Conference on Intelligent Robots and Systems (IROS)},
  year={2025},
}