In the near future, the robots we see in our homes will likely not be solitary workers, but teams. Just as humans collaborate to move furniture, prepare a meal, or clean a large house, future artificial intelligence systems will need to coordinate their actions to achieve shared goals. This field, known as embodied multi-agent planning, asks a fundamental question: how do we teach machines to work together without getting in each other's way? While recent advances have allowed single robots to understand images and perform simple tasks, the jump to a group of robots operating in the same physical space introduces a new layer of complexity. It is not enough for each robot to know what to do; they must know when to do it, who should do it, and how to hand off objects without causing a collision or a delay.

Researchers have long struggled to measure this ability to cooperate. Existing tests often focus on whether a team eventually finishes a job, treating the final result as the only metric that matters. This approach, however, hides the messy reality of how the work was done. A team might succeed by accident, or they might reach the goal after wasting time, repeating the same actions, or ignoring the rules of the environment. To solve this, a team of scientists from Nanjing University, AgiBot, and Tsinghua University has introduced a new testing ground called CoCoBench. Rather than just asking if a team finished a task, this benchmark dissects the specific ways robots must coordinate, measuring the quality of their teamwork step by step.

The researchers built their test inside a sophisticated computer simulation of a household, a digital environment where virtual robots can open drawers, pick up objects, and move around. They created nearly nine hundred distinct scenarios, each designed to force the robots to rely on one of four specific types of cooperation. The first type is task allocation, where a group must decide how to split up independent jobs, such as having one robot sort cups while another sorts plates. The second is sequential ordering, which requires robots to follow a strict timeline, like opening a cabinet before putting items inside and closing it only after everything is placed. The third is mutual exclusion, a scenario where multiple robots need to use a single shared tool, such as a knife, forcing them to take turns. The fourth is handoff coordination, where one robot must pass an object to another through an intermediate spot, like a table, requiring them to time their movements so the receiver is ready when the item arrives.

For each of these scenarios, the researchers did not simply record whether the robots succeeded. They also measured how well the robots coordinated. They tracked whether the team wasted time by doing the same job twice, whether they violated the necessary order of steps, whether they fought over the shared tool, or whether they left each other waiting. This allowed them to separate a successful outcome from a well-coordinated process. A team could finish the task but still receive a low score if they did so by ignoring the rules or by working inefficiently.

When the researchers tested eleven of the most advanced artificial intelligence models available today, the results revealed a surprising truth about machine teamwork. The models that performed best overall did not necessarily excel at every type of cooperation. Some models were excellent at splitting up tasks but struggled badly with taking turns on a shared tool. Others were great at following a sequence of steps but failed when they had to pass an object to a partner. This suggests that the ability to coordinate is not a single, general skill that a robot either has or does not have; instead, it is a collection of distinct abilities that must be developed separately.

The study also examined how the robots communicated. When the researchers gave the robots a central planner that could see everything and direct the team, the results were significantly better than when the robots had to make decisions based only on what they could see themselves. Adding a simple communication channel helped the independent robots, but it did not fully close the gap with the central planner. This indicates that the main difficulty for these systems is not seeing the world, but rather planning the group's actions logically. In fact, when the researchers removed the visual images from the test and gave the robots only text descriptions of the situation, their performance did not drop much. This suggests that the bottleneck is not in recognizing objects, but in the high-level logic of managing a team.

As the researchers increased the number of robots in a team from two to three and then to four, the difficulty of the task grew. The success rate for the teams declined as more agents were added, particularly for the smaller, less powerful models. This shows that adding more workers does not automatically make a job easier; it increases the complexity of the coordination required. The researchers found that while the most advanced models remained relatively stable, the weaker models struggled significantly as the team size grew, often failing to complete the planning within the allowed time or breaking the rules of the simulation.

Perhaps the most critical finding was that looking only at whether a task was completed is misleading. The researchers discovered that some models achieved a high success rate by taking shortcuts that violated the rules of coordination, such as grabbing a tool while another robot was still using it, or placing an object before the container was open. These teams reached the goal, but they did so through chaotic and illegal behavior. By introducing a score that measured the legality and quality of the coordination, the researchers showed that a team can "win" the game while playing it poorly. This distinction is vital for developing robots that can work safely and efficiently in real homes, where following the rules of interaction is just as important as finishing the chore.

The work presented in this paper does not claim to have solved the problem of robot teamwork. Instead, it provides a much clearer way to see where current systems succeed and where they fail. By breaking down coordination into specific, measurable parts, the researchers have shown that building a team of robots requires more than just making individual agents smarter. It requires designing systems that can handle the specific demands of sharing space, time, and tools. As we move toward a future where machines work alongside us, understanding these nuances of cooperation will be essential to ensuring they do so without confusion or conflict.