[[TOC]]
The added complexity of dual or multi-arm manipulation presents many challenges that may not be present in the single manipulator case. This higher complexity means that dual arm manipulation requires more advanced system integration, high level planning and reasoning, as well as viable control approaches.
- Similarity to operator
- Flexibility and stiffness
- Manipulability
- Cognitive motivation 探索类人物理交互与认知的关系
- Human form factor 据称类人型台的机器人性能更好
- non-coordinated manipulation 双臂执行不同任务
- coordinated manipulation 双臂执行同一任务的不同部分
- goal-coordinated manipulation 手臂之间没有物理交互，类似双手打字
- bimanual manipulation 存在物理交互 （图1）
第一大类与单臂无异；第二大类才是本文主题
Domestic
- folding laundry
- providing attendant care and kitchen support for the elderly
- performs cooperative manipulation task with a human, or transports the object autonomously
Industrial
- parts assembly
- crew-assisting robotic system for space exploration
本节介绍有closed chain systems特性的文献，例如the grasp matrix, the internal forces, the restraint properties, the force/load distribution, the inherent redundancy as well as the implications of manipulation of deformable sheets and the use of mobile manipulators in dual arm manipulation
根据多臂末端与物体的交互方式，可以分为两种协作操作任务，这也是双臂机器人建模的主要影响因素：
- Cooperative robot manipulators holding an object with fixed Agrasp points (Fig. 2(a)) 即机械臂与物体无相对运动
- Cooperative robot manipulators holding an object by means of contact points or contact areas (Fig. 2(b)) 即机械臂与物体可以相对运动
第一类指的是具有固定抓握的双臂操作的建模 (Fig. 2(a))。而第二个主要是指通过手指操纵对象 (Fig. 2(a), 1(c))。
Grasp Matrix 描述了接触点或抓取点处的速度和力与映射到物体质心的速度和力之间的动静力学（kineto-static）关系。两个手持物体的机械手的Grasp Matrix（图2(a)）将广义力从附着在抓取点的参考系 {ci} 映射到附着在物体上固定点的参考系 {o}（通常是质心）可以定义如下：
其中，$p_{oci}$ 是参考系{ci}和{o}的相对位置，$O_{3}, I_{3}$代表$3\times 3$的零矩阵和单位矩阵，$S(\cdot)$是一个斜对称矩阵，用于生成叉积。在接触点（接触区域）的情况下，接触建模对于导出抓取矩阵至关重要；在这种情况下，由（1）给出的 G 可以表征完整的抓取矩阵，而包含通过接触传递的力/扭矩分量的选择矩阵产生抓取矩阵。
接触建模确定了末端执行器扳手在接触点处将到达物体的分量，从而确定了传递到物体的速度分量。有以下三种主要类型的接触模型：
- Frictionless point contact
- Frictional point contact or hard finger model
- Soft contact or soft finger model
接触方式对第二种协作操作任务，即机械臂与物体存在相对运动，在建模时候的影响更大。上述抓取形式及接触类型（在非固定抓取的情况下）用于导出控制操作任务的约束。
trajectory tracking
- input–output linearization
- Hybrid force/position control
- impedance control
- sliding mode control, robust adaptive control as well as neuro-adaptive control
- Intelligent control using neural networks and fuzzy systems
regulation
visual servoing
视觉伺服在控制回路中的参与方式可以分为
- image-based 直接使用观测与理想位置的2D图像特征之间的差异作为控制反馈
- position-based 使用物体被观测到的3D位姿与理想位姿之间的差异作为控制反馈（该位姿从RGB/D图像估计得到）
- hybrid methods
此外，相机的安装位置也有几种：
- End-effector mounted 装在机械臂末端
- Fixed in the workspace 固定在工作空间的某个位置
- Active head 装在机器人头上（常见于家庭中的拟人机器人）
motion, trajectory and manipulation planning
- Randomized potential field planning techniques with the introduction of transit (where the arms move independently and the object remains static) and transfer (where the arms and the object form a closed kinematic chain) subpaths
- Randomized Path Planner algorithm
- RRT
- programming by demonstration (PbD)
- Hidden Markov Models and Gaussian Mixture Regression
- combines Dynamical Systems movement control with a Programming by Demonstration approach
- learning a policy which maps the state to a control vector
- integration of elements from systems theory with tools from cognitive methodologies
- Is it meaningful to mimic the (proven successful) configuration of biological systems, such as humans? Should more advanced sensors be used to cover deficiencies in planning, modelling, and control, or should the latter be developed to cover deficiencies in sensors?
- high-level reasoning / scene understanding