The emergence of motor synergy and energy efficiency during whole-body locomotion learning has garnered attention as researchers explore how complex movements can be simplified. A recent study led by G. Li and M. Hayashibe sheds light on this topic, employing deep reinforcement learning (DRL) to investigate how locomotor synergies develop as agents learn to walk and run.
Muscle synergies, the basic modules coordinating complex muscle activities across many joints, are foundational to human and animal movement. While previous studies primarily focused on upper limb synergies, this research explores whether these patterns also occur within the high-dimensional whole-body locomotion systems of humanoid robots.
Using the advanced simulation capabilities of the Isaac Gym engine, the researchers trained humanoid agents with 28 degrees of freedom (DoFs) on various walking and running tasks. Interestingly, the study found synergy emerged without explicitly encoding it within the learning framework, driven by the agents’ responses to their environment.
By varying the weight of symmetry constraints during training, the findings indicated a compelling relationship between symmetry, energy efficiency, and motor synergy. Higher symmetry weights corresponded with enhanced energy efficiency and more coordinated movement patterns. The data revealed, “The synergy emerged during the learning of walking and running tasks, illustrating the correlation between motor synergy, energy efficiency, and gait symmetry.”
This aligns with the hypothesis posited by previous studies, which suggest the natural evolution of symmetric and synergistic gaits as mechanisms promoting energy conservation during movement. Levels of synergy increased alongside agents’ proficiency and coordination as they moved from random patterns to more efficient, human-like gaits.
Through systematic training, distinct patterns of motor synergies incompletely mirrored human locomotion behaviors. The agents exhibited greater efficiency with more symmetric motions, as evidenced by enhancements seen across several metrics monitored during the trials, including average speed and cost of transport. “Even before considering the symmetry aspect, synergy emergence could be observed,” the authors pointed out, emphasizing the inherent efficiency of the learned behaviors.
According to the researchers, fostering synergy within machine learning opens new avenues for enhancing robotic mobility and control strategies, as well as offering insights applicable to human motor learning. They noted, “The humanoid agent spontaneously learned synergistic gaits during the learning process, similar to humans and animals.” This observation suggests future research could probe the interactions between genetic and experiential influences on motor synergies, potentially revolutionizing how robotic systems mimic human locomotion.
By integrating concepts of motor synergy with DRL, the study not only provides fascinating results but reinforces the potential for creating advanced humanoid systems capable of natural movements. Future work may address how developing locomotor synergies can be optimized within various gait challenges, paving the path for more sophisticated robotics and novel therapeutics for human mobility.