H-GAP: Humanoid Control with a Generalist Planner (2312.02682v1)

Abstract: Humanoid control is an important research challenge offering avenues for integration into human-centric infrastructures and enabling physics-driven humanoid animations. The daunting challenges in this field stem from the difficulty of optimizing in high-dimensional action spaces and the instability introduced by the bipedal morphology of humanoids. However, the extensive collection of human motion-captured data and the derived datasets of humanoid trajectories, such as MoCapAct, paves the way to tackle these challenges. In this context, we present Humanoid Generalist Autoencoding Planner (H-GAP), a state-action trajectory generative model trained on humanoid trajectories derived from human motion-captured data, capable of adeptly handling downstream control tasks with Model Predictive Control (MPC). For 56 degrees of freedom humanoid, we empirically demonstrate that H-GAP learns to represent and generate a wide range of motor behaviours. Further, without any learning from online interactions, it can also flexibly transfer these behaviors to solve novel downstream control tasks via planning. Notably, H-GAP excels established MPC baselines that have access to the ground truth dynamics model, and is superior or comparable to offline RL methods trained for individual tasks. Finally, we do a series of empirical studies on the scaling properties of H-GAP, showing the potential for performance gains via additional data but not computing. Code and videos are available at https://ycxuyingchen.github.io/hgap/.

• The paper introduces a general-purpose planning framework that leverages MoCap trajectories and MPC to achieve superior humanoid control.
• It utilizes a large-scale MoCapAct dataset to train a generative model that replicates human motor behaviors across various control tasks.
• Scaling experiments indicate that while larger models boost motion imitation accuracy, dataset diversity is crucial for optimizing downstream performance.

Introduction to Humanoid Control Challenges

Humanoid control is a critical area of research with promising applications ranging from integration into human-centric environments to creating realistic computer-generated animations. This field, however, poses complex challenges due to the intricate optimization required to navigate the high-dimensional action spaces that characterize humanoid control systems. Often, data derived from human motion capture (MoCap) offer a valuable resource, aiding in the optimization process and bringing human-like finesse to the resulting models.

Generalist Approach to Humanoid Planning

The Humanoid Generalist Autoencoding Planner (H-GAP) introduces a novel approach to this challenge, utilizing a generative model trained on a large-scale dataset of MoCap-derived state-action trajectories. Unlike existing methods that may need further online interactions or cater to specialized tasks, H-GAP is equipped to learn from an offline dataset—MoCapAct—without requiring additional interactions post-training. Further, it can apply the acquired knowledge to new control tasks by leveraging a planning method known as Model Predictive Control (MPC), showcasing its flexibility and ability to generalize.

Comparative Performance and Empirical Insights

Empirical studies showcase H-GAP's ability to accurately represent and generate human motor behaviors learnt from the dataset. When deployed in a variety of downstream control tasks, H-GAP has demonstrated comparable or superior performance to existing offline reinforcement learning methods that train separate, specialized policies for each task. Significantly, H-GAP even surpasses traditional Model Predictive Control (MPC) strategies that utilize the actual physics model, emphasizing the strength and robustness of the learned latent action space and action prior in H-GAP.

Scaling and Future Directions

An exploration into the scalability of H-GAP reveals noteworthy findings: while increased model size improves the accuracy of motion imitation, larger models don't guarantee better performance in downstream control tasks. This could be due to a decrease in the diversity of generated samples with larger models. Additionally, when it comes to dataset size, larger and more diverse training sets contribute to better performances, suggesting that more expansive human MoCap datasets can further propel advancements in humanoid control models. This research can inspire subsequent developments in methods for humanoid control that are effective, scalable, and tailored for a diverse array of applications.