We propose the expert composer policy, a framework to reliably expand the skill repertoire of quadruped agents. The composer policy links pair of experts via transitions to a sampled target state, allowing experts to be composed sequentially. Each expert specializes in a single skill, such as a locomotion gait or a jumping motion. Instead of a hierarchical or mixture-of-experts architecture, we train a single composer policy in an independent process that is not conditioned on the other expert policies. By reusing the same composer policy, our approach enables adding new experts without affecting existing ones, enabling incremental repertoire expansion and preserving original motion quality. We measured the transition success rate of 72 transition pairs and achieved an average success rate of 99.99%, which is over 10% higher than the baseline random approach, and outperforms other state-of-the-art methods. Using domain randomization during training we ensure a successful transfer to the real world, where we achieve an average transition success rate of 97.22% (N=360) in our experiments.

Reinforcement learning (RL) policies are prone to high-frequency oscillations, especially undesirable when deploying to hardware in the real-world. In this paper, we identify, categorize, and compare methods from the literature that aim to mitigate high-frequency oscillations in deep RL. We define two broad classes: loss regularization and architectural methods. At their core, these methods incentivize learning a smooth mapping, such that nearby states in the input space produce nearby actions in the output space. We present benchmarks in terms of policy performance and control smoothness on traditional RL environments from the Gymnasium and a complex manipulation task, as well as three robotics locomotion tasks that include deployment and evaluation with real-world hardware. Finally, we also propose hybrid methods that combine elements from both loss regularization and architectural methods. We find that the best-performing hybrid outperforms other methods, and improves control smoothness by 26.8% over the baseline, with a worst-case performance degradation of just 2.8%.

This paper proposes the transition-net, a robust transition strategy that expands the versatility of robot locomotion in the real-world setting. To this end, we start by distributing the complexity of different gaits into dedicated locomotion policies applicable to real-world robots. Next, we expand the versatility of the robot by unifying the policies with robust transitions into a single coherent meta-controller by examining the latent state representations. Our approach enables the robot to iteratively expand its skill repertoire and robustly transition between any policy pair in a library. In our framework, adding new skills does not introduce any process that alters the previously learned skills. Moreover, training of a locomotion policy takes less than an hour with a single consumer GPU. Our approach is effective in the real-world and achieves a 19% higher average success rate for the most challenging transition pairs in our experiments compared to existing approaches.