TABLE 3.

Summary of publications on learning-based multi-robot control.

Reference	Tasks	Method/architecture	Robot model	Policy input	# of robots trained (tested)	Scalability	Needs localization
Agarwal et al. (2020)	Coverage, line, Formation	RL/GNN	Point mass	Absolute pose	5 (2–10)	Yes	Yes
Li et al. (2022)	Path planning	RL + LfD/CNN + FC	Holonomic	LiDAR velocity, position	3–5 (3–5)	No	Yes
Yan et al. (2022)	Formation + path planning	RL/RNN	Ackermann steering	Distance angle	3–5 (3–5)	No	No
Blumenkamp et al. (2022)	Path planning	RL/GNN	Holonomic	Absolute pose	5 (5)	−	Yes
Li et al. (2020)	Path planning	LfD/CNN + GNN	Point mass	Binary map	4–12 (4–14)	Yes	No
Tolstaya et al. (2020a)	Flocking	LfD/GNN	Point mass	Absolute pose	100 (50–150)	Yes	Yes
This paper	Triangular formation	LfD/CNN + GNN	Nonholonomic	LiDAR	5 (3–9)	Yes	No

“−” represents a case in which no result was presented.