Arnold: A Multi-Task, Multi-Embodiment Muscle Transformer Policy

Researchers have introduced Arnold, a transformer-based policy for musculoskeletal control that can excel in various tasks and forms, as outlined in a paper on arXiv (2508.18066v2). This study tackles the complex issue of managing high-dimensional, nonlinear musculoskeletal models of the human body, a critical scientific challenge. While earlier advancements in machine learning have led to in-silico policies that perform well in specific skills such as reaching, object manipulation, and locomotion, these agents are limited to single tasks. In contrast, Arnold utilizes a combination of behavior cloning and reinforcement learning to manage 14 demanding control tasks, including dexterous object manipulation and locomotion, achieving performance that matches or surpasses that of specialist policies. A notable feature of Arnold is its sensorimotor vocabulary, which provides a compositional representation of diverse sensory modalities, allowing the policy to generalize effectively. The paper was released as a replace-cross type and can be accessed at https://arxiv.org/abs/2508.18066.