Towards biomimetic control of robotic or paralyzed limbs

我们的长期目标是从生物运动中学习，为人工系统（机器人，假肢）和辅助人类控制（FES，神经接口）设计有效和新颖的控制策略。迄今为止，我们在平行的次级目标中侧重于两个方面： 1-新的分析工具适用于生物运动系统的非线性参数领域，切换混合策略和强加的拓扑结构的困难的背景下。在NSERC的资助下，我们开发了现在由NSERC/I2 I支持的软件创新：创建切换系统中响应模式的自动分类和识别工具，以估计每个模式的动态，尽管切换特性。他们的目标是分发给神经科学家，神经耳科医生和临床医生在该领域。 2-基于脑干和脊髓中神经回路的特殊性质（拓扑结构）的眼-头和眼反射生物控制策略的新模型（由CIHR资助）。这些模型改变了行为和神经层面上对数据的解释，并提出了新的协议来测试运动系统在更自然的混合感觉环境中的完整性，如在日常生活中。这是现在唯一可能的，因为我们的算法可以处理不同的条件。最近，对手臂伸展的探索（由NSERC资助）预测，我们在眼头协调中使用的控制策略与灵长类动物手臂或腿的轨迹以及脊髓的拓扑结构是一致的。其结果是比机器人文献中通常使用的控制简单得多-即无需规划的移动。 所有这些系统在空间组织和感觉运动融合方面都使用类似的网络拓扑结构。这表明一个可能的一般控制理论的所有平台，无论是他们堆叠和旋转，或节段肢体。因此，长期目标是： 形成运动系统的仿生控制策略，允许执行简单任务而无需优先轨迹规划 作为短期目标，将对生物学中发现的上述特征进行评估，并将其正式化用于一般应用，其中： · 一种控制器结构或拓扑结构，类似于脑干和脊髓中的神经连接（对称性，传感器-运动相互作用的位点），以首先在空间上1D然后3D地嵌入具有模式选择标准（开关，映射矩阵的阶）的动态模式。 · 传感器和目标平台之间以及平台之间的最佳非线性增益场，它决定了轨迹的动力学和曲率。 · 通过执行中央控制（类似于皮层和小脑）进行控制器“调整”的学习策略，并允许独立调整运动速度和轨迹，而无需重新计算轨迹计划。 主要假设是：任务错误直接发送到所有参与平台（没有单独的目标）;每个段和终点的轨迹作为网络动态特性而演变，而不是像经典机器人技术中那样通过计算预先施加。如果这可以推广并容易地调整为不同的机械系统，它将在几个领域产生影响：尝试驱动瘫痪肢体的接口的“自然感觉”，如果来自替代肌肉的输入可以保持其自然激活模式，则使用假肢的学习曲线更快，以及用于远程设备操作的更自然的接口系统。目标是让用户保持他/她的习惯激活模式，而不是开发可能与他们的神经功能相冲突的新模式。最后，这应该会导致智能机器人系统能够自动适应环境，计算需求低。

Our long-term objectives are to learn from biological movement and devise efficient and novel control strategies for artificial systems (robots, prostheses) and for assisted human control (FES, neural interfaces). To date, we have focused on two streams in parallel sub-objectives: 1- New analytical tools applicable to biological movement systems in the difficult context of non-linear parametric fields, switched hybrid strategies and imposed topologies. Funded by NSERC in the past, we produced software innovations now supported by NSERC/ I2I: to create automated classification of response modes in switched systems and identification tools to estimate the dynamics of each mode despite switching characteristics. They are targeted for distribution to Neuroscientists, Neuro-otologists and Clinicians in the field. 2- Novel models of biological control strategies for eye-head and ocular reflexes (funded by CIHR) based on the special nature (topology) of neural circuits in the brainstem and spinal cord. These models change the interpretation of data at both behavioural and neural levels, and suggest new protocols to test the integrity of motor systems in more natural mixed-sensory environments, as in daily life. This is only possible now because our algorithms can handle diverse conditions. More recently, explorations on arm reaching (funded by NSERC) predict that our control strategies used in eye-head coordination are consistent with primate arm or leg trajectories and the topology of the spinal cord. The result is much simpler control than typically used in the robotic literature – i.e. movement without planning. All of these systems use network topologies similar in their spatial organization and in the use of sensorimotor fusion. This suggests a possible general control theory for all platforms, be they stacked and rotatory, or segmental limbs. So the long-term goal is: TO FORMALIZE A BIOMIMETIC CONTROL STRATEGY FOR MOVING SYSTEMS THAT WILL ALLOW EXECUTION OF SIMPLE TASKS WITHOUT A-PRIORI TRAJECTORY PLANNING As short term-objectives, the mentioned characteristics found in biology will be evaluated and formalized for general applications, with: • A controller structure or topology analoguous to neural connections in the brainstem and spinal cord (symmetry, sites of sensor-motor interactions) to imbed dynamic modes with mode selection criteria (switch, order of mapping matrix) first spatially 1D then 3D. • Optimal Non-linear gain fields between sensors and their target platforms, and between platforms, which determine the dynamics and curvature of trajectories. • Learning strategies for controller “tuning” by executive central control (analogs cortex and cerebellum) and allow independent adjustments of movement speed and trajectory, without the need to re-compute a trajectory plan. The main assumptions are: task error sent directly to all participating platforms (no separate goals); trajectories for each segment and the end-point evolve as a property of the dynamics of the network, rather than pre-imposed by computation as in classical robotics. If this can be generalized and easily tuned for different mechanical systems, it would have impact in several areas: a 'natural feel' for interfaces that attempt to drive paralyzed limbs, a faster learning curve for the use of artificial limbs if the input from alternate muscles can preserve their natural activation patterns, and a more natural interface system for remote device operation. The goal is to allow the user to keep his/her habitual activation patterns, rather than develop new ones that may be in conflict with their neural capabilities. Finally, this should lead to smart robotic systems that adjust to contexts autonomously, with low computational demand.