00:00Opening a fridge, sorting waste and slicing vegetables.
00:04These are simple tasks for humans.
00:07For robots, however, they are signs of a much bigger breakthrough.
00:10One Chinese startup believes humanoid robots are nearing a point
00:14where they can take on a much wider range of tasks.
00:16But how close are they really?
00:18We're at the Beijing offices of embodied AI startup Spirit AI.
00:22It was established in 2024 and is one of China's fastest growing
00:26and most advanced companies developing embodied AI foundation models
00:30for software models that are powering robot brains.
00:33So behind me you can see their Moz1 wheeled humanoid robot
00:37and we have this mocked living room scenario.
00:40So right now it's taking some items of waste,
00:43so a banana peel, bits of cardboard, bottles,
00:48and putting them in a bin and putting bottles in a fridge.
00:52So it has to, you know, chain together lots of movements.
00:54So these are called long horizon tasks
00:56and it's what a lot of humanoid software developers are building
01:01to develop the next generation of more powerful robot brains.
01:04Behind every movement is a vast training effort.
01:07During our visit, dozens of people fitted with sensors
01:10repeated everyday actions like opening fridges,
01:13picking up bottles, and cutting vegetables.
01:15Their movements are recorded and used to help train robot brains
01:19to interact with the physical world.
01:21Spirit AI says it has about 1000 people doing such work
01:25and has collected hundreds of thousands of hours of training data that way.
01:29The company says that has helped its humanoids achieve a 90% success rate
01:33for simple tasks in structured living room environments.
01:37As you can see behind me, there's been quite considerable strides in progress
01:41in the technology powering these robot brains, but there's still a long way to go.
01:45The co-founder and chief scientist of Spirit AI,
01:48Gao Yang, just told us in an interview that he expects it would take at least eight years
01:52for humanoid robots to safely enter people's homes
01:55and complete the domestic tasks like you can see this robot doing behind me.
02:01I think that in the near term, more robot applications will be in industrial settings,
02:06and that will likely be the first wave of adoption.
02:11Industrial environments are more structured,
02:13and the actions required are relatively repetitive.
02:17If you can get a robot to perform one action well,
02:20you can place it into a production environment
02:23where it can replace people in doing a lot of labour.
02:27I think the next one to two years will be the initial window
02:30for robots to take off in industrial applications.
02:35Gao draws parallels with the development of AI and says bots are nearing a key moment.
02:42We believe that around the middle of 2027, we will reach what we call a GPT 3.0 moment.
02:50This may sound a bit technical, but it is an analogy with large AI models.
02:57When OpenAI released its models, it went through a series of versions,
03:01and GPT 3.0 is the most representative example.
03:05At that stage, you can say almost anything to the model,
03:08and it can give you a fairly reasonable response.
03:11Although the response is not always exactly what you wanted,
03:14it is still a relatively reasonable one.
03:18For now, Gao said there are still hurdles to tackle.
03:21Tasks requiring precise hand movements, like unscrewing a bottle cap, remain difficult,
03:27and robots can still struggle when faced with situations they haven't encountered before.
03:31As these machines become more capable, however,
03:34the question of how to keep them safe for humans will loom ever larger.
03:38Spirit AI says its robots are programmed to stop automatically
03:41if they detect excessive forces in any interaction.
03:44But if humanoid robots eventually move beyond factories and enter homes,
03:48making them safe, may be just as big a challenge as making them smart.
Comments