11 EigengraspsTop9 Grasp Planning10 The Dynamics Engine


10 The Dynamics Engine

GraspIt! has two main modes of operation. In the first one, which has been assumed in all of the previous chapters, the user directly interacts with the simulation world by moving objects around or moving robot joint draggers. The collision detection system disallows any movement that would bring two bodies in collision, and stops any such movement at the point of contact. We refer to this operation mode as "static operation".

In real life however, objects move around as a result of contact and external forces. In GraspIt! these phenomena are taken into account when using the dynamics engine. This engine computes the contact forces that prevent interpenetration and the velocities and accelerations of all the bodies in response to contact forces, joint constraints and external forces. The resulting motion of the bodies is integrated over time in a time-stepping procedure resulting in a full simulation of a grasping process (as opposed to just assessing the quality of a final grasping posture, as we do in static operation).

10.1 User interaction in dynamics mode

The key difference between static and dynamic operation is in the way the user interacts with the simulation world. In statics, the user directly moves bodies around, and sets robot joint values. None of these are allowed in statics. The only way the user can affect the simulation world is by setting desired values for the robot DOF's. Built-in controllers for the robot DOF's take care of the rest, simulating true dynamic operation. DOF controllers generate motor forces; as a result of these robot joints move and potentially touch other objects. Bodies are also affected by gravity, and will fall off the world if they are not supported by an obstacle (such as a floor or table). Obstacles are only part of the dynamics engine in the sense that they provide contacts for other bodies. They never move in dynamics.

Use the Dynamics start button in the toolbar to start or pause the dynamics engine. While the engine is running, you should see the dynamics timer advancing. During this time, you will see bodies move as a result of gravity, or DOF motor forces if the desired DOF values have been set. For a quick demo, see the Getting started chapter of this manual.

Currently, the only way for a user to specify a desired position for a robot in dynamics mode is to use the AutoGrasp feature. Use the Grasp -> Auto Grasp menu just as you would in statics. When dynamics are on, this will set the robot desired DOF values at either the DOF max or min, depending on the default velocity specified in the robot configuration file. In the future, we intend to provide an interface for a user to set desired robot DOF values during dynamics execution.

10.2 DOF controllers

The simulated DOF controllers have proven to be very difficult to calibrate and tune. Currently, we are implementing PD controllers and attempt to generate trajectories with smooth acceleration and velocity profiles from current DOF values to desired DOF values. See the DOF::PDPositionController(...) function for details. These controllers are very sensitive to the proportional and derivative gains used, and we have no better method of setting these than empirical tuning. Even like this, for complex dynamic simulations, we often see erratic behavior. We are hoping to improve the DOF control mechanism in the future.

10.3 Implementation

For all the details on the dynamics engine implementation, see the Publications chapter. Here we provide just a quick overview.

The dynamics engine works in time steps. At each time step, two main tasks are performed:

The core of the dynamics engine, and also its most important computational effort, is the LCP solver. We have implemented a version of Lemke's algorithm for this, found in the dynamics.cpp file.

10.4 Under development: dynamics improvements

Unfortunately, the dynamics engine exhibits occasional instability, which we would like to address in the future. These might be caused by the LCP solver, the time step integration, or both. It has been suggested by experts in the field that the LCP solver is the primary suspect. We are hoping at some point to replace the current solver with a better one, based on newer methods proposed in the literature.

We would also like to cut down on computation time by providing the LCP with a "warm start" using the solution from the previous time step. This is made more difficult by the fact that contacts are always broken and then re-computed between time steps, making it harder to keep track of previous time step solutions. We have put in place a mechanism for "inheriting" contact parameters between time steps, you will find this in the Contact class. However, with the current LCP solver, we have not been able to obtain an improvement using this information. Contact "inheritance" though has been left in place (although disabled for now) in the hope that future LCP solvers might be able to take advantage of it.

Copyright (C) 2002-2009 Columbia University

11 EigengraspsTop9 Grasp Planning10 The Dynamics EngineContents