A reinforcement learning approach to multi-robot planar construction

Strickland, Caroline (2019) A reinforcement learning approach to multi-robot planar construction. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (20MB)

Abstract

We consider the problem of shape formation in a decentralised swarm of robots trained using a subfield of machine learning called reinforcement learning. Shapes are formed from ambient objects which are pushed into a desired pattern. The shape is specified using a projected scalar field that the robots can locally sample. This scalar field plays a similar role to the pheromone gradients used by social insects such as ants and termites to guide the construction of their sophisticated nests. The overall approach is inspired by the previously developed orbital construction algorithm. Reinforcement learning allows one or more agents to learn the best action to take in a given situation by interacting with their environment and learning a mapping from states to actions. Such systems are well-suited to robotics, as robots often interact with complex environments through a variety of sensors and actuators. When reinforcement learning is applied to a multi-agent system, it is called 'multi-agent reinforcement learning' (MARL). The main feature that MARL offers is flexibility | a multi-agent decentralised system can have agents added, removed, or reconstructed without need for rewriting the system. This allows for more robust solutions due to its ability to cope with failure. With the use of simulators paired with MARL, we can effectively learn policies that result in the formation of unique shapes. This is a vast improvement over hand-coded solutions, as it removes dependence on hard-coded actions. Reinforcement learning eliminates the need for writing control algorithms in the first place | which tend to be be extremely task-specific and time-consuming.

Actions (login required)

View Item