top of page

Dual-arm manipulation

 

 

Introduction

 

 

  For the last fifty years, industrial single-arm manipulators have been widely used in factory environments. In contrast, dual-arm systems are just emerging to be regarded as mainstream in robotics. The dual-arm system significantly increases object manipulability, and in the human environment there indeed exists many tasks requiring coordination between two arms. Examples include washing dishes, putting luggage in a compartment or trunk, stirring milk into coffee, etc. One of the benefits to using the dual-arm system is the capability of manipulating bigger objects owing to the intrinsic high degrees of freedom (DOF) for task-space redundancy. In addition, the closed kinematic chain formed by the dual-arm system and the object usually increases the object-grasping stability. 

  The force interaction between the two arms and the object is unavoidable in dual-arm manipulation. Force control is commonly used as the main control strategy. In general, the force control can be categorized into two groups: indirect control and direct control. Indirect control accomplishes force control by means of motion control. One of most common control algorithms is the impedance control. Thus, many researchers focus on the impedance behavior of the manipulator. The robot DLR uses virtual spatial springs,connected to the end-effectors in Cartesian space, and a coupling spring, connected to arms, to accomplish compliant behavior. Researchers at KIST uses virtual spring-dampers connected to every joint and proposes a virtual dynamic model to reduce the impact during object manipulation.

  Besides compliant behaviors, some researchers deal with planning and control. Based on work by Yul et al. , researchers at KIST transmit human motions to a humanoid robot with a motion-capture system to manipulate objects . Researchers at JPL combine both vision and kinesthetic information to track both manipulator and object. Shuji et al. use three tactile sensors equipped at the end-effector to detect the object surface orientation and keep the hand direction normal to the object surface in three dimensions.

Force control

 

 

  Here, instead of directly focusing on a sophisticated manipulation task, we try to establish a straightforward control strategy that is inspired by a human manipulating a large object with his two arms. In addition to the algorithm, a dual-arm system is constructed for experimental evaluation. From a human perspective, three factors that affect the object manipulation after grasping are sense of space, sense of touch, and vision. The vision is important for grasping, but may be minor after the object is held. As a result, the former two factors are addressed in this paper. We investigate master-slave control strategy in three levels with different complexity. The right arm is equipped with a force/torque sensor (attached on the palm) and is considered as the slave side. The left arm without sensory feedback is considered the master side. The first-level controller is an open-loop position controller using spatial relation. Once the left arm is commanded to move, the right arm follows the left hand and maintains a certain spatial relation. Without using force information, the right arm can’t adjust its force to hold the object when slippage happens. Thus, the second-level controller introduces compliance on the slave side by relating the force error to the velocity state. Two abilities are required for the palm to stably grasp the object: orienting the surface of the palm with that of the object and keeping the desired normal force between the palm and the object. The slave side with these two abilities can generate a fine touch of the object if the spatial movement is small. When the spatial movement is large, the force information on the arm is not sufficient to decouple the hand orientation or hand displacement. Thus, the motion of the master side is a priori for dual-arm spatial manipulation with the object. To remedy this limitation, we devised a third-level controller whose control input is created by fusing the position and control input and velocity control input of the first two levels with a Kalman filter infrastructure. Thus, the final controller is capable of simultaneously controlling both position and normal force. The slave side can regulate its pre-defined trajectory under external force disturbance by keeping its desired normal force. The proposed three-level control strategy is inspired by human dual-arm manipulation and provides a good connection between control law and the physical world.

Three stage control strategy

 

 

Open loop control

Open loop control

When the object is grasped, the object length, L, using the right hand position,PR,and the left hand position,PL,can be computed

Surface Normal control

Surface Normal control

The concept of the surface normal control is mapping the torque to the object rotation by using PID compensation structure.

The virtural resultant force

The virtural resultant force

The desired normal force and the modified force form a virtual resultant force. The virtual resultant force maps to the displacement using a PID controller

Normal Force control

Normal Force control

The normal force control uses a similar principle, but the force/torque measurement and compensation motion generation are in different directions.

Hybrid control structure

Hybrid control structure

The user determines the position and orientation of the left hand. By using the left-hand command and the object length, the open-loop control calculates the right-hand position and orientation. By using the measured force/torque, the compliant control calculates the change position and orientation. These two control inputs are fused by the Kalman filter. The fused position and orientation are used to control the slave side.

Kalman filter structure

Kalman filter structure

In the hybrid control, the desired trajectory is analogous to the right-hand position and orientation, which is calculated by the open-loop position control. The velocity change is analogous to the change position and orientation, which is calculated by the compliant control.

bottom of page