GRASPER: A Permissive Planning Robot

GRASPER: A Permissive Planning Robot*

Scott Bennett

SRA Corporation 2000 15th Street North Arlington, VA 22201 bennett@sra.com (703) 558-7508

Execution of classical plans in the real world can be problematic. Small discrepancies between a planner's internal representations and the real world are unavoid­ able. These can conspire to cause real-world failure even though the planner is sound and, therefore, "proves" that a sequence of actions achieves the desired goal. The difficulty, of course, is that the planner's proof is con­ tingent on its internal world model precisely capturing all relevant features of the external real world. This is seldom the case, particularly in robotics where uncer­ tainties abound. Small but unavoidable sensor errors preclude accurate knowledge of the state of the world . Worse, the planner's own characterization of the effects of its actions are themselves only approximations. Real­ world execution of a sequence of actions can introduce and quickly magnify inconsistencies with the internal mi­ cro world.

We have been investigating one response to this diffi­ culty called permissive planning [Bennett, 1993; DeJong and Bennett, 1993; 1995], a machine learning extension to classical planning. Our video presents GRASPER, a permissive planning robotic system that learns to ro­ bustly pick up novel objects. In permissive planning, machine learning techniques are employed to refine the planning algorithm based upon empirical observations of success and failure in the world. Alteration of the plan­ ner is accomplished through planner bias adjustment. Planner bias is an inescapable facet of classical planning. It refers to the preference that a classical planner exhibits when it produces one particular plan from among the (of­ ten very large) set of distinct plans it could in principle construct. In the GRASPER system, planner bias is ad­ justed through Explanation-Based Learning of schemata which eclipse the native bias inherent in the conventional searching planner employed as the explanation engine.

The GRASPER system consists of a six degree of free­ dom Prab RTX scara-type manipulator, an overhead camera, a frame grabber, and an IBM RT computer run­ ning Lucid Common Lisp. GRASPER constructs plans to lift plastic pieces of a children's puzzle. No a pri­ ori models of the objects are given to the system. The

*This work was conducted at the Beckman Institute at the University of Illinois at Urbana-Champaign. Funding was provided through the Office of Naval Research under grant NOOOl 4-91-J 1563.

Gerald Dejong

University of Illinois/Urbana 405 North Mathews Urbana, IL 61801

dejong@cs.uiuc.edu (217) 333-0491

pixel patterns from the camera are converted into sim­ ple polygons which serve as the object representations to the planner. These representations (like any internal representation of a real physical object) are flawed , being only approximations to the actual objects. GRASPER is given a conventional classical axiomatization for how ob­ jects can be surrounded , how closing the gripper applies a friction force between the fingers and the object, how sufficient friction establishes a grasp, how a grasped ob­ ject can be manipulated, etc. The axiomatization, like the represented sensory data, captures the real world only approximately: The represented coefficient of fric­ tion is not precise. Operators that represent arm move­ ment effects only approximate the motions they claim to perform. The forces that the robot fingers apply to a surrounded object also only approximate the operator's effects, and so on.

Using its initial knowledge, a plan is constructed to lift a designated object . Not surprisingly, the initial plan usually fails in the real world. Following the principle of permissive planning, the planning process, rather than the represented domain theory, is blamed for the short­ coming. The planning bias is adjusted. Through bias adjustment over several failures, the real-world effects of produced plans are made to conform to the projection of the original action sequence. The final pi ck-up schema can be interpreted as 1) squeezing harder than the world knowledge claims is necessary, 2) selecting grasp points along faces which are more nearly parallel, 3) select­ ing grasp points closer to the object 's center of geom­ etry than believed necessary, and 4) opening the gripper wider than believed necessary while approaching the tar­ get object .

A unique feature of permissive planning is that the planning algorithm rather than the underlying incorrect representations are adjusted to overcome execution fail­ ures. This is the dual of the more conventional approach which successively debugs the planner's domain knowl­ edge, leaving the planning algorithm unchanged . As refined operator definitions and object representations become more elaborate, the complexity of the planning process can grow dramati cally. One advantageofpermis­ sive planning is that the operator definitions and object representations remain as the implementor originally de­ fined them.

Permissive planning is not conceived as a general solution to all of the difficulties that arise in planning under uncertainty. However, it offers a unique set of advantages that can complement other approaches.