Abstract Despite the growing demand for capturing diverse motions, collecting climbing motion data remains challenging due to difficulties in tracking obscured markers and scanning climbing structures. Additionally, preparing varied routes further adds to the complexities of the data collection process. To address these challenges, this paper introduces a physics‐based climbing controller for synthesizing climbing motions. The proposed method consists of two learning stages. In the first stage, a hanging policy is trained to naturally grasp holds. This policy is then used to generate a dataset containing hold positions, postures, and grip states, forming favourable initial poses. In the second stage, a climbing policy is trained using this dataset to perform actual climbing movements. The episode begins in a state close to the reference climbing motion, enabling the exploration of more natural climbing style states. This policy enables the character to reach the target position while utilizing its limbs more evenly. The experiments demonstrate that the proposed method effectively identifies good climbing postures and enhances limb coordination across environments with varying slopes and hold patterns.

Since the introduction of DeepMimic [Peng et al. 2018a], subsequent research has focused on expanding the repertoire of simulated motions across various scenarios. In this study, we propose an alternative approach for this goal, a deep reinforcement learning method based on the simulation of a single-rigid-body character. Using the centroidal dynamics model (CDM) to express the full-body character as a single rigid body (SRB) and training a policy to track a reference motion, we can obtain a policy that is capable of adapting to various unobserved environmental changes and controller transitions without requiring any additional learning. Due to the reduced dimension of state and action space, the learning process is sample-efficient. The final full-body motion is kinematically generated in a physically plausible way, based on the state of the simulated SRB character. The SRB simulation is formulated as a quadratic programming (QP) problem, and the policy outputs an action that allows the SRB character to follow the reference motion. We demonstrate that our policy, efficiently trained within 30 minutes on an ultraportable laptop, has the ability to cope with environments that have not been experienced during learning, such as running on uneven terrain or pushing a box, and transitions between learned policies, without any additional learning.