Robot Weightlifting and Motor Synergies
	Loosely based on Olympic weightlifting, the three-link manipulator at left learns to exploit dynamics and lift weights far greater than possible given the static torque constraints. Joint and contact constraints break traditional control engineering solutions for acrobot-like problems, and three underpowered, but actuated joints also make typical machine learning solutions impractical. The approach used in this work makes use of a hierarchical control structure that incorporates a number of elements found in human motor learning: imitation, shaping, knowledge of results (reinforcement), and motor programs.
The main result of this project involves not the success or failure of a particular learning algorithm, but rather the structure of the control policies before and after learning. In particular, coordinated movement is related to motor "synergies" that actively couple individual joint motions by way of the robot's control system. Moreover, without such coupling the robot becomes considerably impaired and unable to lift all but the lightest payloads. This is true for both the simulations and the real robot pictured below.

Related Publications
Learning to Exploit Dynamics for Robot Motor Coordination
	Ph.D. thesis, University of Massachusetts Amherst, 2003. Committee: Andrew G. Barto (chair), Neil E. Berthier, Andrew H. Fagg, Roderic A. Grupen and Richard E. A. Van Emmerik. [pdf]
Velocity-dependent dynamic manipulability
	M.T. Rosenstein and R.A. Grupen. In Proceedings of the IEEE International Conference on Robotics and Automation, vol. 3, 2424-2429, 2002. [pdf] [ps.gz]
Robot weightlifting by direct policy search
	M.T. Rosenstein and A.G. Barto. In Proceedings of the Seventeenth International Joint Conference on Artificial Intelligence, vol. 2, 839-844, 2001. [pdf] [ps.gz]
A robotic weightlifter that learns to exploit dynamics
	M.T. Rosenstein and A.G. Barto. In Studies in Perception and Action VI: Eleventh International Conference on Perception and Action, 25-28, 2001.
From elementary movements to coordination for a robotic weightlifter
	M.T. Rosenstein and A.G. Barto. In Abstracts of the Third International Symposium on Progress in Motor Control: From Basic Science to Applications, p. 40, 2001.

Movies
File	Description	Download
std2_NoLearn	Standard solution for the real robot before learning.	2.2 MB
std2_Learn	Standard solution for the real robot after learning.	1.9 MB
wtliftsim	Three-minute animation of various simulation results.	18.8 MB
Note: To view the movies, use a QuickTime player (Mac/Win), MPlayer (Mac/ Linux), or at least version 2.80 of XAnim (Unix).

updated 28-Nov-2003 mtr@cs.umass.edu