中国科学院自动化研究所机构知识库

传统的串联刚性机器人凭借其高速度、高精度和高稳定性的优势能够代替人完成许多危险、繁重和重复性的任务，在国防工业和国民经济中发挥了重要作用。但是，面对日益增长的社会需求，例如代替人实现精密操作，与人在同一工作空间内安全地协作，适应非结构化的动态环境等，传统机器人还存在局限性。与传统机器人相比，肌肉骨骼机器人通过模拟人体的骨骼，关节和肌肉结构，以及肌肉和关节间的驱动方式，从结构上提供了更好的灵活性、可靠性、柔顺性、安全性和适应性。因此，肌肉骨骼机器人是提高现有机器人性能和满足社会需求的一种潜在途径。但是，肌肉骨骼机器人的高冗余、强耦合和强非线性也给控制带了巨大挑战。现有的肌肉骨骼机器人运动控制方法在缺乏大量监督样本、高维状态反馈、即时和稠密奖励信号的情况下，仍存在运动精度低，泛化能力弱和缺乏多任务连续学习能力等问题。由于人类和动物可以灵活、柔顺地控制肌肉骨骼系统，与此相关的神经机制成为了实现肌肉骨骼机器人高性能控制的一种直接参考。因此，本文通过分析，提炼和借鉴相关神经机制，提出了一系列脑启发式运动控制方法，主要内容和贡献如下：

1. 针对运动精度低和泛化能力弱的问题，本工作受生物体的时变肌肉协同激活机制启发，综合利用阶段性和强直性肌肉协同激活模式，提出了新型的时变肌肉协同激活计算模型和神经肌肉控制算法。所提方法在复杂肌肉骨骼系统上开展了仿真验证。实验表明，该方法有效地提高了肌肉骨骼系统运动学习的速度，精度和泛化能力。

2. 针对显式设计肌肉协同激活模式的局限性，本工作受运动皮层的神经编码机制启发，设计了仿运动皮层神经编码特性的循环神经网络（Recurrent Neural Network, RNN）和肌肉骨骼机器人运动学习方法。仿真实验表明，在无需显式设计肌肉协同激活模式的情况下，本工作仍能实现肌肉骨骼机器人和冗余机械臂的精准运动学习和泛化，简化了算法设计的难度。同时，所设计的RNN能够展现与运动皮层类似的一致集群响应和稳定、抗干扰特性，产生具有协同激活模式的肌肉控制信号，有较好的生物可解释性。

3. 针对肌肉骨骼机器人在不同场景下进行连续自主学习的需要，本工作受运动皮层中神经流形和多任务学习机制的启发，构建了能够表征RNN所习得知识和技能的神经线性子空间，提出了受神经线性子空间调控的肌肉骨骼机器人多任务连续学习方法。仿真实验表明，所提方法能够在不改变网络结构和不显式利用历史样本进行训练的情况下，实现在不同运动范围和不同重力环境下的多任务连续学习。

4. 考虑到灵巧、柔顺的肌肉骨骼机器人在操作任务方面具有独特优势，本工作针对肌肉骨骼机器人的操作任务，提出了一种基于环境吸引域策略和时不变肌肉协同激活机制的仿人操作和控制方法。所提方法在肌肉骨骼系统的轴孔装配任务上开展了仿真验证。结果表明，所提方法能够利用肌肉骨骼系统的柔顺性实现高于控制精度的操作，而且可在部分肌肉驱动器损坏的情况下鲁棒地完成操作任务。

综上所述，本论文针对在缺乏大量监督样本、高维状态反馈、即时和稠密奖励的情况下所存在的肌肉骨骼机器人控制问题，设计了一系列脑启发式运动控制方法，提高了肌肉骨骼机器人运动学习的精度、泛化能力、多任务连续学习能力和柔顺操作能力，对于肌肉骨骼机器人的运动控制，机器人和神经科学的交叉融合具有积极意义。

With advantages of high speed, high precision and high stability, traditional serial and rigid robots have replaced human in many dangerous, heavy and repetitive tasks, which plays an important role in national defense and economy. However, in the face of growing demands for substituting human in high precision manipulation, cooperating with human safely within the same working space, and adapting to unstructured environments, traditional robots still have limitations. Compared with traditional robots, the musculoskeletal robot can provide better flexibility, reliability, compliance, safety and adaptability with the imitation of human skeleton, joint and muscle structure, and the driving mode between muscles and joints. Therefore, musculoskeletal robot is a potential way to improve the performance of existing robots and meet the growing demands. However, the high redundancy, strong coupling and strong nonlinearity of the musculoskeletal robot also bring many challenges to control. Under the circumstances without a large number of supervised samples, feedback of high-dimensional states, and immediate and dense rewards, existing control methods still have bottlenecks in the movement accuracy, generalization and multi-task continual learning. As human and animals can control the musculoskeletal system flexibly, their related neural mechanisms are the most relevant references for realizing the high performance control of the musculoskeletal robot. Therefore, with the analysis, selection and imitation of related neural mechanisms, this thesis proposes a series of brain-inspired control methods as below:

1. The first work aims to improve the movement accuracy and generalization. Inspired by the biological mechanism of time-varying muscle synergy, this work proposes a computational model of time-varying muscle synergy and a neuromuscular control method with the utilization of phasic and tonic muscle synergies. The proposed method has been verified by the control of a sophisticated musculoskeletal system in simulation. The results indicate that the method improves the speed, accuracy and generalization of motion learning.

2. The second work aims to overcome the limitation of establishing the model of muscle synergy explicitly. Inspired by the mechanism of neural coding in primate motor cortex, this work designs a recurrent neural network with motor-cortex-like characteristics and proposes a motion learning method of the musculoskeletal robot. The simulation results indicate that the method can realize precise motion learning and generalization without establishing the model of muscle synergy explicitly. Furthermore, the method also has great biological plausibility. Specifically, the RNN demonstrates motor-cortex-like consistent population response and stability, and generates muscle commands with co-activation patterns.

3. The third work aims to realize the continual learning of musculoskeletal robot under different situations. Inspired by the mechanism of neural manifold and multi-task learning in primate motor cortex, this work establishes the linear neural subspace to represent the learned knowledge and skill of RNN, and proposes the multi-task continual learning with the modulation of linear neural subspace. The simulation results suggest that without the change of network structure and the explicit utilization of history samples, the method can realize the continual learning of a musculoskeletal robot in multiple tasks with different movement scopes and gravities.

4. The fourth work aims to realize human-like manipulation through utilizing the unique advantage of flexible and compliant musculoskeletal robot in manipulation task. Based on the strategy of attractive region in environment and the mechanism of time-invariant muscle synergy, this work proposes a human-like manipulation and control method. The method has been verified by the peg-in-hole assembly task of a sophisticated musculoskeletal system in simulation. The results show that the proposed method can not only realize high-precision manipulation with the compliance of musculoskeletal system, but also achieve the assembly task under the situation of part muscles are destroyed.

In a summary, for the control problems in the circumstances without a large number of supervised samples, feedback of high-dimensional states, and immediate and dense rewards, this thesis designs a series of brain inspired motion control methods for musculoskeletal robots. The proposed methods improve the accuracy and generalization of motion learning, multi-tasks continual learning ability, and compliant manipulation ability, which has positive significance for motion control of musculoskeletal robots and integration of robotics and neuroscience.