Why Don’t We Have Better Robots Yet? | Ken Goldberg | TED

Folding the laundry at 3-6 folds per hour is way higher than the 0 folds per hour I am currently doing.

For manipulating objects with our hands, we have a set of sensors in our fingers that robots can only dream of. Pressure, temperature, texture, weight, slipperiness... I couldn't pick up a coffee cup if vision was my only sense.

Watching this, I realized my cluttered house might actually be a strategic move to keep robots at bay!

I'm here because I work at an Amazon warehouse five minutes from home by car. Just checking out the competition

"upload dates" don't cut it. The dates of the actual talk is what matters the most, with technology advancing as quickly as it is.

Robots don't need to be perfect, they just need to cover 3 criteria. 1) graceful failure states. A robo vaccuum sounds great, but if it is going to choke on a string, or smear a mess and make things worse rather than better... Yeah, no thanks. If a folding robot rips a shirt a week then I'm not keeping it around. I feel like this is the biggest hurtle right now. The price of many things has gone down, and reliability gone up... But those rare fails are big fails that undo all of their savings. They don't need to be perfect. They just need to fail better. 2) they need to save time. I keep looking at robot mowers, especially during spring allergies. My fear is that I will spend more time picking one, and programming it, and maintaining it, and replacing it than I will save by just mowing the lawn myself. The argument that learning the skills to maintain it are more valuable than the skill of mowing the lawn is not lost on me. But what I need in life is time savings, not a new hobby. And the savings needs to be well beyond 10:1. It doesn't matter if a robot is significantly slower than me at a task, as long as it saves me time. A robo mower may take 20-40 min a day to cover the same yard I do in an hour a week... But if I don't have to do anything for that hour a week, then it is still a net benefit to me. 3) it has to be a money saver. So many appliance style robots come with a high up front cost, and a yearly subscription, and generally do a worse job than I do at the task. Like time, it has to be in the 10:1 savings range over the life of the product to make it worth it. So either a loss leader with a subscription like printers have moved towards, or a high up front cost for years of little to no maintenance like traditional appliances. Robot mowers that cost thousands, and then cost monthly are never going to break even before they are replaced. It may free up some time so I can work more... But me working for my robot kinda misses the point of the hired help. 😅 But I think we are getting close. Robo vaccuum cleaners are already just about there. And we are on the cusp of many others. I think the big mistake companies are making is designing specific use robots, or building humanoid robots, when neither are particularly easy or cost effective. Something that can use the equipment I already have, but without the fall hazard of a bipedal robot is ideal. Something on rollers or treads, with arms and a mast for cameras is the simple answer to so many of the issues. Then as long as it can push a vaccuum, or manipulate laundry, or reach the dishes and stove... We can add capabilities with processing and software upgrades over time. But the actual form factor and mechanic side of things is there already. Just need to dress it up in a way people will enjoy, and gain more skills.

11:23 That kid has 5 fingers on his right hand (not including a thumb, which might be out of view). And that Rubik's Cube is 3x4x2.

Please always add the date of the talk, thanks!

I was a mechanical engineer at UC Berkeley/Lawrence Berkeley Laboratory for 30 years. I’ve built automation for production lines. I seen Goldberg’s laundry folding robots over the years. I’ve witnessed the progress in technology and software. When he says be patient, the robots are coming he means not in our lifetime for robots of the “Jetson’s” sense in our homes. Expect continuing incremental progress for commercial applications.

6:55 It's not that people "don't like doing this work", it's that the companies providing the work are mistreating and underpaying them.

I honestly don't think we'll make real progress in making humanoid robots like this until we can give them tactile sensors. Like, we get so much subconscious information about our surroundings from our skin (well also the tiny hairs on our skin). I feel like even just knowing "I am touching something" without having to "see" it would help so much with clumsiness.

The key is continuous learning. I'm still "learning" how to make my morning omelette. Each day I get new feedback; try new techniques, gather more information and learn from tiny mistakes. Robots must do the same.

Good talk. But it almost seemed like there was a "laugh track" everytime he said something.

This is precisely what I want to see in robotics; efforts to automate away drudge work. Kudos to them.

This is the type of presentation I can never forget. Knowledge is shared with humour. Well done Kim!

that reminds me of the shirt-folding board that reduces a significant amount of time with a dollar. We always need to keep the problem-solving mindset while approaching the problem. I am impressed by how persistent the honourable researchers are with getting rid of chores for everybody. Applause the real heroes!

At 2:25 he’s basically showing you Roberto from futurama 😂😂

some current issues: scale & targeted specificity of training (not very general beyond task, lack gen int) inference can be slow inference can be expensive inferences via foundation models are disjointed --- with foundation models / LLMs that will help knockout the scale/generalization issue, looks like some companies are using hierarchical methods to make progress ie figure01 costs per inference seems to be declining rapidly speed of tokens per second or inferences a second is also making progress. Disjointed nature/statelessness of inference is the last hurdle; research is currently investigate State models & memory units to learn from past activations to inform current inference. Though prob not 100% necessary, as tasks can be broken up into many smaller tasks, this would help with coherence esp of goals across inference and fluidity -instead of each inference popping into existence at that moment not aware of the context, 'thoughts', why of previous inference and then handed the chat history up to that point to seem somewhat continuous. Though there are hacks like having it think aloud it's plans and goals needed and currently doing but some nuance is lost thus not ideal.

I'm so glad to hear an insider's view on the ongoing challenges of modern robotics. ❣️

“Don’t believe robots are clumsy? Here’s proof… I got one to cut and style my hair.”