51
0Episode: Robots Learn to Move: How Xiaomi Is Teaching Machines to Think
Duration: approximately 7 minutes
Level: B1 (Intermediate)
---
[Mike]: Hey everyone! Welcome back to "Learn English with Podcasts." I'm Mike.
[Sarah]: And I'm Sarah! Today we're talking about something really cool — robots that can learn to move and do things in the real world.
[Mike]: That sounds like science fiction, but it's actually happening right now. Xiaomi just released a project called Xiaomi-Robotics-1. It's a foundation model for robots.
[Narrator]: Foundation model 是一种大型预训练模型，可以应用于多种任务。类似 GPT 在语言领域的角色，机器人基础模型也能处理各种操控任务。
[Sarah]: A foundation model for robots? What does that mean exactly?
[Mike]: Think of it like this. In language AI, we have models like ChatGPT that can do many things — answer questions, write stories, translate languages. A robot foundation model is similar, but for physical actions.
[Sarah]: So instead of learning to write text, the robot learns to move objects, open doors, fold clothes?
[Mike]: Exactly. And what makes this special is the data. Xiaomi trained this model on over 100,000 hours of real-world robot data. That's a huge amount.
[Narrator]: Pre-training 指模型在大量数据上进行初步训练，学习通用能力，之后再针对具体任务做微调。
[Sarah]: 100,000 hours? That's like working 24 hours a day for over 11 years!
[Mike]: Right. And it covers more than 1,700 different scenarios — homes, offices, factories, even outdoor spaces.
[Sarah]: How did they collect so much data? Did they have thousands of robots working all day?
[Mike]: Sort of. They used something called embodiment-free data. It's captured with a special device called UMI — a handheld robot controller. A human operator moves the controller, and the robot records the motion.
[Narrator]: Embodiment-free 意味着数据不绑定特定机器人形态，可以用于训练不同类型的机器人。
[Sarah]: Oh interesting. So the data isn't from one specific robot, it's more general. That makes sense for a foundation model.
[Mike]: Exactly. Then in a second stage, they use real robot data to teach the model how to actually control physical robots. This is called post-training.
[Sarah]: What kind of tasks can the robot do after training?
[Mike]: All sorts of things. Tidying a sofa, sorting shoes in a cabinet, putting kitchen items away, packing suitcases. It can even do things like refill a printer or load laundry into a washing machine.
[Sarah]: That's impressive. And how well does it work?
[Mike]: Really well. When they gave it less than 10 hours of training per new task, it already achieved a 75% success rate. With 40 hours of training, it reached 85%.
[Narrator]: Success rate 是成功率，表示任务完成的百分比。75% 意味着每 100 次尝试有 75 次成功。
[Sarah]: And compared to other robot models?
[Mike]: It beats the competition. On simulation benchmarks — that's like virtual tests — Xiaomi-Robotics-1 scored the highest on all four major tests. One test showed a 58% improvement over the second-best model.
[Sarah]: Why does having more data and a bigger model help so much? Is there something special about how they trained it?
[Mike]: Yes. They found something called scaling behavior. As the model gets bigger and sees more data, it gets consistently better. The improvement doesn't slow down — it keeps going up.
[Narrator]: Scaling behavior 指模型性能随数据和参数规模增长而稳定提升的现象，是大模型的关键特性。
[Sarah]: That's encouraging. It means there's still a lot of room to improve.
[Mike]: And here's what's really practical about this. The model can learn new tasks with very little data. If you want it to learn a brand new task, you only need a few hours of demonstrations.
[Sarah]: Demonstrations meaning someone shows the robot how to do the task?
[Mike]: Right. The human performs the task while the robot watches and records. Then the model learns from those examples. It's like teaching a child by showing them how to do something.
[Sarah]: So what does this mean for the future? Will we all have robot helpers at home?
[Mike]: We're getting closer. This kind of foundation model is a big step. Instead of programming each robot task separately, you train one general model that can adapt to many situations.
[Sarah]: It's like the difference between learning one recipe and learning how to cook in general. A general cooking skill lets you handle any recipe.
[Mike]: Great analogy. And because the model scales well, as more data becomes available and computers get more powerful, these robots will keep getting better.
[Sarah]: I'm excited about this. It feels like we're entering a new era where robots can actually be useful in everyday life.
[Mike]: Me too. Well, before we wrap up, let's review today's vocabulary.
[Sarah]: First, "foundation model." A foundation model is a large AI model trained on lots of data that can be adapted to many different tasks.
[Mike]: "Pre-training" is the first stage of training a model on a huge dataset to learn general skills.
[Sarah]: "Post-training" comes after pre-training. It fine-tunes the model for specific tasks or environments.
[Mike]: "Scaling behavior" means that as a model gets bigger and uses more data, its performance keeps improving.
[Sarah]: "Embodiment-free" describes data that isn't tied to one specific robot. It can be used to train different types of robots.
[Mike]: "Demonstration" means showing someone how to do something by doing it yourself.
[Sarah]: Great job today, everyone! Thanks for listening to "Learn English with Podcasts." See you next time!
[Mike]: Bye-bye!
