Build a multi-agent ROS transcript

CYRUS WONG: What if your command lines were not just for Git commit and file system. What if it is the nervous system for a robot swarm? I received an email from Google Cloud. The challenge, architecture with a resilient multi-agent orchestrator that translates high-level intent into synchronized, low-level execution. Let's do it. [MUSIC PLAYING] Today, we are pushing Angular CLI past the chatbot. We are turning it into a multi-agent orchestrator for six humanoids.

No heavy ROS middleware, no hardcoded control plane, just the high-level intents and emergent coordination. Let me explain it a bit background, existing source code here. Here is the source code enzyme. It's just messy I don't want to have to investigate the detail here, but my target is that I just want Antigravity CLI to help me understand the source code and then do what I want. I just want to control it and do the action and investigate the system.

OK, since we have input the username and password for it and the IP address, it is able to get a lot into the robot. And now I want to know about, is this really, really they got the information of the robot? I asked one more question. How is the battery level of robot 1? You see, the battery level here. If it is fully charged, it will be 12.6. But now, the robot still, OK, it got 11.1 volts and it is not in low battery condition.

Let me try to ask the robot to do something. Can you tell me what robot 1 can do for me? Here is the action we were engineering. They found it. The robot can do some actions. For example, they can ask it to go forward, turn right, turn left, and then do push up, do sit up. Maybe we can also do some high level instruction. For example, create an exercise in four steps, just like that. Starting my first step workout. OK, we confirmed that Anti-gravity CLI is able to control the robot.

I suspect that the robot action is not really completed. I want Anti-gravity CLI to go into the robot, and then drill down all of the action, go down the action list. As I suspect, there are more actions the robot might be able to do. Oh, nice. It's newly discovered some action for me, I really don't know. For example, this is the so-called wave dance, wings up, seal dance, no mad. Oh, this is some action. OK, let's try it.

Let me do. The left, right, dab for robot 1. Great. OK, one more trial. I don't know what is wings up. Now do wings up and say something. I believe I can fly. Preparing for takeoff. Woo-hoo! Oh, this is the first time I see it doing this action. OK, it's time to do the multi-agent demonstration. And this is a bit more upfront. We are not just running one instance or one robot or agent. And we can have a swarm of six robots together using the subagent concept to do the cooperation, coordination, something high-level instruction, so that they can do it together.

Now, I instruct Anti-gravity CLI to create some subagent to control robot. Now, this is my problem. Treat each robot as a subagent and then perform a wave starting from leftmost robot to the right one. Hi. Nice. They completed the action. And here's the report as the layout is telling me that oh, the robot is the-- who is the row 1, who is in row 2, and then the leftmost, what is the timing delay for them to say hi to us.

OK, just typing the comment is a bit tedious and not really user friendly. How about we just directly talk into the Anti-gravity CLI, and then Anti-gravity will be able to run a program, it is able to allow us to capture my voice and then answer me by voice? Start a voice chat. OK, here is the voice chat. Hello, can you help me ask a robot to say hi to me? Hello, we are so excited to say hi to you. Hello. Now, ask all robot go forward.

Walking forward. Oh, they are walking forward. What we just saw is shipped in developer tools. We are moving from writing commands to managing agents. But this has a lot left to solve. Observability is here. How do we debug when agent number 4 decides it not want to cooperate? How do we scale this to 60 robots, 600 robots, and so on? We are bridging the gap between prototype and embody system using tools already using everyday.

If you want to see the code and the prompt schema, check out the link below. Thank you for watching, and have fun building. [MUSIC PLAYING]