An Epistemological Investigation into Learning of Behaviour Trees

Abstract

Automated robotic action planning is a key problem in AI. Using an autonomous forklift as an example, we apply cognitive science-inspired approaches to improve the accomplishment of complex tasks in the context of utility machines [1]. Such tasks can look like the following: a forklift truck is forced to lift a pallet, which is blocked by another. One solution is to move the blocking pallet first to access the cleared targeted one. However, a solution should not have to be trained manually for each occurrence, but rather achieved autonomously, through a reactive action planning method, capable to cope with dynamic changes in the environment. For safety reasons, the overt behaviour resulting from any such deployed algorithm needs to be “readable” by humans.

To achieve such comprehensible machine cognition, a modification of a Behaviour Tree (BT) is to be learned through Genetic Programming, which is inspired by Evolutional Biology and 4E Cognition [2], [3]. An epistemological investigation is concerned with the concepts of knowledge representation and creation in the context of BTs. We aim to explicate, which definitions and considerations are of practical utility in the context of the development of AI applications. By adding this philosophical perspective, the problem of action planning in AI becomes an interdisciplinary investigation.

Three implementation steps are conducted to demonstrate the advantages of the cognitive science-informed approach. In the first stage, we implement and empirically assess modifications of BTs in the benchmark example of the Pac-Man game. We then combine the lessons learned with an analysis of the actual targeted deployment environment to select a suitable variant. Next, we compare the selected method to a BT hand-crafted by experts in a simulation of the forklift truck scenario. Finally, we aim to deploy the method on the actual forklift truck, to demonstrate its usability in the real world.

The impact of this master’s thesis is bidirectional, reflecting its interdisciplinary approach. On the one hand, a cognitive science-informed method is presented for learning BT-based action planning in AI. On the other hand, the philosophical aspects underlying “knowledge” in machine cognition are made explicit for the example application scenario.

References

[1] AIT Austrian Institute of Technology GmbH, "HOPPER - AIT Austrian Institute Of Technology", ait.ac.at, Sep. 30, 2022. https://www.ait.ac.at/en/ueber-das-ait/center/center-for-vision-automation-control/complex-dynamical-systems/hopper?no_cache=1 (accessed Apr. 18, 2023).

[2] M. Colledanchise and P. Ögren, "Behavior Trees in Robotics and AI: An Introduction," CoRR, vol.  abs/1709.00084, 2017. Available: http://arxiv.org/abs/1709.00084 (accessed Apr. 18, 2023)

[3] N. Nilsson, "Teleo-Reactive Programs for Agent Control," arXiv, Dec. 31, 1993. doi:10.48550/arXiv.cs/9401101.