Videos from research into robust contact methods for robotics
This video demonstrates the effectiveness of one of my contact methods for the problem of a simulated manipulator grasping a tennis ball. This task is notoriously difficult to simulate, and few contact methods are capable of doing so; my method does so in O(n3), which is faster than any other known method.
The only other example of physically simulated grasping is Miller's excellent GraspIt! software; that software uses a contact method specially developed for grasping, while my method is general purpose (grasping, locomotion, and more). Miller's method exhibits worst-case exponential time complexity and expected-case O(n4) time in the number of contact points.
Videos from my advanced penalty method for rigid body simulation
Reduction of oscillation
This video shows how the advanced penalty method is able to avoid the oscillations that plague the standard penalty method. The cube is placed such that it penetrates the ground plane, and then the penalty method is applied. Notice that the cube is pressed straight upward, along the direction of the contact normal.
Before the penalty method is applied.
A short time after the penalty method is applied.
Ability to handle piles of objects
Piles and stacks of objects are problematic for most contact methods. This video shows a stable state of a pile of objects of different shapes and masses inside a box. As can be seen from the video, the objects are resting without movement.
Videos from research into the Task Matrix
Robots used in the videos
The simulated robots used in videos. Note that the layouts of the degrees-of-freedom are quite different; for example, the simulated Asimo employs five DOF in the arm while the mannequin uses six. Additionally, the heights of the two robots are quite different.
Notes on Task Matrix videos
 The robots in these videos are kinematically simulated. As a result, locomotion is depicted in an unsophisticated manner (sliding across the floor rather than walking). Note, however, that an implementation on a physically simulated or true humanoid robot will not require any porting to get the task programs using locomotion to function properly; locomotion in the Task Matrix is accomplished abstractly by giving planar position and orientation commands to the robot base.
 The movement shown in these videos is often not hominine. This is a result of our implementation of the underlying algorithms that drive the robot. It is possible to produce more human-like movement by altering these inverse kinematics and motion-planning algorithms, without affecting the task programs in any way.
 The mannequin has no fingers. Thus, we simulate an electromagnetic mechanism for grasping objects; the objects are affixed directly to the mannequin's hand. However, this mechanism makes for uninteresting videos of the grasp and release task programs.