This project contains a version of blocks world that is formulated for hierarchical task decomposition. It involves three levels of problem spaces. There is sufficient evaluation knowledge so that there is no search/uncertainty at every level. The top level has a single operator: move-block, which moves a block (moving-block) to a destination. The destination can be the top of another block or the table.

The next level consists of two operators: pick-up and put-down and they arise in an operator no-change for move-block. The rules for these are found under move-block. This level introduces the gripper, which can be holding a block or empty and is a structure on the top state.

The bottom level has a variety of operators: open-gripper, close-gripper, move-gripper-above, move-gripper-down, and move-gripper-up. The structures manipulated by these operators are part of the gripper structure on the top state. These operators arise in an operator no-change for both pick-up and put-down and the rules for them are under pick-up.

Includes gripper status to state. Execution is done through top-state rules that simulate changes to the io-link

Works with chunking, which compiles the actions in the substates into rules that apply at the top-state. Use the command "learn -e" to turn on chunking, run the agent once, init-soar then run the agent again to see how the learned productions eliminate the need to subgoal in the second run.

Soar capabilities
  • Hierarchical task composition via subgoaling
  • Internally simulates external environment including an i/o link
  • Can learn procedural knowledge (enable with 'learn always')
Download LinksExternal Environment
  • None.
Default Rules
  • simple.soar
Associated PublicationsDeveloper
  • John Laird
Soar Versions
  • Soar 8,9
Project Type
  • VisualSoar