- code and example
- see Hannaford et al., 2016
Behavior trees (BTs) emerged from video game development as a graphical language for modeling intelligent agent behavior. However as initially implemented, behavior trees are static plans. This paper adds to recent literature exploring the ability of BTs to adapt to their success or failure in achieving tasks. The “Selector” node of a BT tries alternative strategies (its children) and returns success only if all of its children return failure. This paper studies several means by which Selector nodes can learn from experience, in particular, learn conditional probabilities of success based on sensor information, and modify the execution order based on the learned information. Furthermore, a “Greedy Selector” is studied which only tries the child having the highest success probability. Simulation results indicate significantly increased task performance, especially when frequentist probability estimate is conditioned on sensor information. The Greedy selector was ineffective unless it was preceded by a period of training in which all children were exercised.
To run the demo, as used in the ArXiv paper above,
> python2 bt_icra_sim.py ARG
Where ARG is one of ['S0', 'S1', 'S2', 'S2g', 'S3', 'S4', 'S5' ] where
| code | Selector Type |
|---|---|
| S0 | Normal "dumb" selector |
| S1 | Tick in descending order of probability of success |
| S2 | P(S|F) condition P(S) on the environment state |
| S2g | Same as S2 with Greedy selector |
| S3 | Tick in descending order of cost |
| S4 | Tick in descending order of utility |
| S5 | Tick in descending order of utility conditioned on env. state |
A new demo has been created to illustrate a concrete robot control scenario capable of integration with a real robot arm target (e.g. Raven or dVRK).
- To experiment without a ROS/CRTK Installation: use
abt_move_cp_example.pyand refer to details. - To experiment WITH a full ROS/CRTK installation, use
crtk_abt_move_cp_example.pyand follow setup directions in its comments.
This is the official python version of the Behavior3 library, originally written in Javascript.