Hacker News Reader

Note that a mouse does not need to solve the problem from scratch like a computer does. They are born with a general solution to the problem of movement in their brain, which has been arrived at by evolution.

To replicate this, you can't expect to get away with only replicating the complexity of the mouse. You potentially need a computer with as much complexity as the evolutionary algorithm which led to the mouse's movement algorithm.

Note that a mouse does not need to solve the problem from scratch like a computer does. They are born with a general solution to the problem of movement in their brain, which has been arrived at by evolution.

Only for some cases. I've watched a horse being born and seen pictures of other newborns. The foal can stand within minutes, and the right sequence of moves is clearly built in because it works the first time. Newborn horses can walk within hours and run with the herd within days. Lying down, though, is not pre-stored. That's a confused collapse for the first few days. There's an evolutionary benefit to being able to get up and escape threats early, but smoothly lying down is less important.

This is a faulty argument, that might apply to primitive organisms with similarly primitive movement schemes.

Most complex organisms spend some time inside a protective shell of some form before getting out into the wild, where I suspect they learn to coordinate their orientation sensory organs with their limbs to produce necessary movement patterns.

I agree that animal intelligence and mobility/dexterity is a more useful goal to aim for now than "human-level".

Boston Robotics seems to have proven that one can get very useful and adaptive mobility with existing algorithms and a lot of (apparently manual?) fine-tuning. Power density seems to have played a role there.

I think for manipulation, the first challenge is useful vision and understanding. The robot needs to see the 3d structure and often infer parts that are not visible, or at least be able to scan around to create a complete picture. It needs to identify and/or recall affordances for that type of structure. It also needs to understand the physical properties of the object such as whether it will deform and how much.

For vision the ability to acquire and analyze point clouds now is an advantage over what was generally available in the 80s or 90s. Personally I believe that it is possible to build very useful manipulators using point clouds converted to voxels and then deep learning from that to get affordances and control policies and I think people are doing it.

To be really efficient and effective though, systems that do have inference about the world as a starting point seem more promising. I am reading a book called Surfing Uncertainty by Andy Clark. He seems to be advocating for Hierarchical Predictive Control. Everything is making sense to me. So far I am finding it light on the details of how to actually implement it. I think that Ogma may be doing something like this. It seems like something I need to at least understand.

One thing I think about is to be able to "just" reconstruct 3d oriented surface sections from individual 2d images. Because it seems to me that that is a strong prior to leverage in recognizing whole surfaces or shapes and then from there object parts. There are interesting papers similar this this but so far not quite what I was hoping for. Which in a way might be good because that motivates me to keep learning.

But anyway I think that really strong vision and understanding is foundational to robotic manipulation. And I also think part of the problem people have is that it's so hard that roboticists often give up and deploy very poor vision systems which means they are handicapped before they even started planning movements because they don't have an accurate detailed picture and understanding of what they are looking at.

But also I need to finish Clark's book because I think theories like predictive processing can explain a lot of the advantages that animals have over most robots.

robotic manipulation in unstructured spaces is probably not so hard anymore. when it comes to hardware, the hardware has often been the problem and very flaky, I believe we are at a stage where the software is ready and waiting for the hardware to catch up.

Humans are able to do surgical operations using existing hardware [0], but software can not.

Humans can drive cars on crazy Indian roads, software can not.

Existing hardware is plenty sufficient for manipulating physical world, but we are missing the intelligence part.

[0] https://www.davincisurgery.com/

AI's need special training grounds because they don't have the benefit of evolution. If they don't have realistic bodies in a realistic environment they can't learn to solve the problem of locomotion and manipulation. Have you noticed the explosion of "AI Gyms" in the last few years?

As others have said, we're at a good point with vision and control, but hardware is still expensive, and this slows research. I expect dexterous robots in a decade, BD's already got one dancing better than me.

But, humans can currently learn to teleoperate robots (e.g., surgical robots) much better than AI. So, they are dealing with the same hardware, and it is hardware a priori unfamiliar to the human. Thus, at least in these cases, the difference must be in the learning algorithm.

I think it’s because Atleast in the case of humans a lot of unstructured manipulation is based on learned observations. We can tell by looking how fragile something is. Babies take some time to figure out using their hands fully.

I think we miss that many of the issues faced with getting software-driven systems to accomplish anything in the world is constrained by the interface with the world more than it is by anything computational. Alpha Go can beat the best human at Go, but Alpha Go can't even load itself into memory on a machine that doesn't have ld.so installed and won't execute without the right version of libstdc++ (I'm making up the stack because I have no idea what they actually used, but you get the idea). Human limbs, on the other hand, are a near universal interface to any tactile objects with chemical properties that don't destroy human flesh. The interconnection between the sensors and the CPU units achieves bandwidth and parallelism more comparable to the entire Internet than any single computer attached to a robotic arm.

Besides that, I don't see that computational intractability being solved by efficient heuristic approximation algorithms necessarily helps when those heuristics require data-generating processes that can't be accomplished efficiently. They're more efficient than NP-hard problems that possibly can't be solved before the heat death of the universe, but still, if your decision making function requires data to be trained and that data requires physical experimentation and collection in the world, you're inherently rate-limited by the speed with which experiments can be carried out. That has contributed to stagnation in particle physics, for instance. Physicists aren't using shitty algorithms. It just takes a long time to build a particle accelerator, and you can't do it faster just because you're a computer. Manufacturing processes and collection of physical material isn't made faster by having faster clock speeds in your decision making center.

AI risk enthusiasts tend to hand wave around this by saying a sufficiently powerful AI can learn from simulation instead of experimentation, but surely there are some limits to how scalable this idea is? Simulating Go is trivial. You've got a 361 cell world where each cell can be in one of two states. On the other hand, I just looked up how many atoms are in a human skin cell and I'm seeing 10^14? Trying to simulate human society from first principles of physics to learn optimal war strategy seems like it require more memory than a computer will ever have access to.

>The research question I would pose is, why is robotic manipulation in unstructured spaces so hard? Machine learning has not helped much there. Yet it's a fundamental animal skill. We're missing something that leads to success in that area. Whatever that is, it may be the next thing after machine learning via neural nets.

One of the reasons is simply that human hands are densely covered with mechanoreceptors that transmit a lot of information and are better than sensors for robotic hands.

Right, try doing any sort of fine manipulation while wearing thick rubber gloves and you'll appreciate how hard it is to design robotic hands.

I had a way to avoid that problem. I built a robot end effector that has a 6-DOF force sensor and an end wrench, not a socket wrench. Put it on a low-end $400 robot arm. Got this to talk to ROS. The idea was to have it detect, by feel alone, when the wrench had engaged a bolt head properly. Since all the feel is coming through the wrench part, 6 DOF is enough.

Turned out that a low-end robot arm driven by R/C servos is not smooth enough for this, and not powerful enough to hold up the end effector. Might be worth trying with a better robot arm and stiffer 6 DOF force sensor.

Linear programming is not NP-hard and has a variety of solutions that are polynomial time [0]. It gives me pause when people say some NP-hard problem is easy in the "average" case without talking about what the distribution is.

[0] https://en.wikipedia.org/wiki/Linear_programming#Interior_po...

Integer linear programming is NP-hard, which may have caused the confusion here.

The classic simplex method for real linear programming is exponential-time in worst case, but cubic time in the average case. Newer algorithms can beat that.

I would like to see a good treatment of Lucas like arguments. Most take the form of humans having their own halting problem. But, that objection is completely irrelevant, and it concerns me that objectors do not realize this. Makes me think they are missing something.