BoxLift

Paper: Scalable Multi-Robot Collaboration with Large Language Models: Centralized or Decentralized Systems?
Github: GitHub - yongchao98/multi-agent-framework: LLM multi-agent discussion framework for multi-agent/robot situations.

Key Idea

4 robots-in-a-warehouse tasks with incremental difficulty, Tested on 4 different Archs

The Archs in question:

They differ in how they coordinate and manage information about their plan

Figure from Chen et al. 2024, arXiv:2309.15943. Open full size

But we probably need the tasks more, the authers propose 4 tasks; from their paper:

Task	Robots	Goal	What makes it hard	Collision?
BoxNet1	4, 8, 16, 32 arms	Move each colored box to its goal of the same color. Each arm stays in its cell.	Boxes must pass from cell to cell through many arms.	no
BoxNet2	4, 8, 16, 32 arms	As BoxNet1, but a box moves between cells only through a cell corner.	A corner holds only one box at a time.	yes
Warehouse	4, 6, 8, 10 mobile robots	Bring all boxes to the target region along fixed paths.	Two robots on the same location is a collision. Six actions per robot.	yes
BoxLift	4, 6, 8, 10 robots	Lift every box in the fewest steps.	Heavy boxes need several robots together. The agents see box size but not box weight.	no

From §IV-A and Figure 4. For each task and each robot count, the paper runs 10 trials with different start states. Thus, each task has 40 runs per framework.

Source: Chen, Arkin, Zhang, Roy, Fan (2024), “Scalable Multi-Robot Collaboration with Large Language Models: Centralized or Decentralized Systems?”, arXiv:2309.15943 · Four warehouse-style tasks. Static snapshot.

Results

The HMAS-2 is the most successfull, but not the most efficient one, We can take this as a priority problem. If we told the selector we need less tokens spent, HMAS is a bad choice and should favor the centerlized one. others are irrelevant thu which is gonna make it more or less a binary problem.