5 ms·

What could be the main cause for things being so hard to cross over from simulation to reality? Should we invest some more in stochastic programming to handle

by smokel 2y ago

What could be the main cause for things being so hard to cross over from simulation to reality? Should we invest some more in stochastic programming to handle all the little quirks that cannot be modeled perfectly? Or is the major issue that robots must work together with humans, which are already pretty complex from a biological perspective, let alone from a sociological or historical one.

JohnKemeny 2y ago
Relevant: the mythical non-roboticist (114 comments) https://news.ycombinator.com/item?id=39707356 https://news.ycombinator.com/item?id=39707356
teractiveodular 2y ago
Exception handling. If your web server throws a 500, you have a problem. If your autonomous vehicle throws a 500 and is now blocking a major intersection, everybody has a problem and you can't just hit refresh.
- JumpCrisscross 2y ago
  > If your web server throws a 500, you have a problem. If your autonomous vehicle throws a 500 and is now blocking a major intersection, everybody has a problem and you can't just hit refresh. This isn't unique to robots, e.g. Crowdstrike.
JumpCrisscross 2y ago
> What could be the main cause for things being so hard to cross over from simulation to reality? Sensor noise? (And/or quality?) We're decent at building world models that don't permit impossible behaviour. And we're decent at training agents in those world models. What's unpredictable is how a sensor in the real world will map to a world model.
sgt101 2y ago
Simulations are closed, the real world is open. Weird things happen in the real world. They happen all the time and we expect them to happen and hardly notice when they do. Simulations approximate the model that humans use to reason about the world, but we just ignore all the oddness and can't articulate the oddness until it's made explicit by some failure.
hwillis 2y ago
Human fingertips and skin in general are extremely precise high resolution pressure and vibration sensors that robots almost completely lack. Tactile feedback takes up the second most human sensor bandwidth behind vision. Imagine doing anything without it- interacting with the world through a puppet. The joints in a human arm are direct-drive series elastic actuators with pretty incredible sensitivity. If you were holding a big heavy box and I bounced a ping pong ball off it, you'd probably feel the vibration. That's kind of a big deal. You may not be able to measure absolute weight well, or even relative weight, but you can detect quick changes with amazing precision. Lifting any heavy or large object is inherently wobbly even for a robot, and that sensitivity is important.
- alexey-salmin 2y ago
  This. I often feel it's not appreciated how damn brilliant a human is when you look at it as a robot. Thousands of servos, millions of sensors, high-throughput compute AND all of that self-assembled and self-repaired AND built mostly out of water, carbohydrates and like 5 grams of iron. A meatbag is by far the best robot on the face of earth today, the gap is astonishing. My estimate is that we'll get AGI faster than human-capable robots competing with meatbags cost-wise. We'll have to deploy self-assembling robots throughout the whole supply chain from sand and ore to chips and servos, eliminate human salaries everywhere, only then the price will be comparable to a parent assembling and training a child which was the way for millions of years.
0xf00ff00f 2y ago
Also relevant: https://en.wikipedia.org/wiki/Moravec%27s_paradox https://en.wikipedia.org/wiki/Moravec%27s_paradox
krisoft 2y ago
> What could be the main cause for things being so hard to cross over from simulation to reality? When you are simulating you are leaving out of details. Some knowingly, some without you even being aware of them. Let’s imagine you are working on some perception algorithm for a lidar. You have a cool simulation using geometry to calculate how far each laser goes before they bounce back. Did you remember to implement that the lidar takes some amount of time to sweep around? It is not making instantenous global captures. (Which is a lot easier to code down.) This causes a kind of smearing effect particularly as things move fast relative to the sensor. Do you know if your sensor sweeps with one plane of lasers or multiple? What is the exact geometry? Are you simulating it as the sensor is designed or as it is built? Did you remember that the sun blinds the sensor? And if so, do you simulate it correctly? Including the algorithms it has internally to try to compensate for it? Did you know that retroreflective surfaces interact with the laser differently? In particular if there is moisture on the lidar. The light straying through water droplets cause false readings. Did you know that lidars can, under certain meteorological conditions see the exhaust of diesel vehicles? I hope you planed a fluid dynamics engine into your similation to account for the exhaust plume spreading upwards and expanding. And these are just the niggles of a particular sensor modality from the top of my head. I know similarly long lists for other sensor types. And then we didn’t even got into thermal issues, or the sensor aging or anything. And I bet reality has an even longer list than I do. Is any of this important in your domain? If yes you ignore them at your peril. If no, you will waste your time implementing them in your simulation. But the only way you will know for sure is when you stumble uppon them and they surprise you out of nowhere by breaking your assumptions and your code. > Should we invest some more in stochastic programming to handle all the little quirks that cannot be modeled perfectly? That is robotics. This is what people writing software for robots do. This is what researchers working on robotics problems are working on. If you want a nice and small example of this maybe read up onnKalman filters or the RANSAC algorithm. If you fancy a more academic treatment of the topic I recommend Probabilistic Robotics from Sebastian Thrun. > is the major issue that robots must work together with humans That certainly does not make it easier. Have you seen the video of the russian chess robot breaking its oponent’s finger?[1] That is just the least gruesome of the recent incidents that I can recall. How much is this your “main problem” depens on your problem domain though. 1: https://youtu.be/Mt-1smlom_M?si=19sOPkQf3s6JHoFo https://youtu.be/Mt-1smlom_M?si=19sOPkQf3s6JHoFo
YeGoblynQueenne 2y ago
>> What could be the main cause for things being so hard to cross over from simulation to reality? Resolution. Simulations are too low-resolution and the current dominant approach to learning behaviour (Reinforcement Learning) is too high-resolution (more precisely, too high-variance). So the simulations don't represent reality accurately enough and the learned policies overfit to the simulations' inaccuracies. Learning from pixels sounds like a great idea until you realise the physical world is not quite made of pixels.

by smokel 2y ago

JohnKemeny 2y ago

Relevant: the mythical non-roboticist (114 comments) https://news.ycombinator.com/item?id=39707356 https://news.ycombinator.com/item?id=39707356

teractiveodular 2y ago

Exception handling. If your web server throws a 500, you have a problem. If your autonomous vehicle throws a 500 and is now blocking a major intersection, everybody has a problem and you can't just hit refresh.

JumpCrisscross 2y ago
> If your web server throws a 500, you have a problem. If your autonomous vehicle throws a 500 and is now blocking a major intersection, everybody has a problem and you can't just hit refresh. This isn't unique to robots, e.g. Crowdstrike.

JumpCrisscross 2y ago

> What could be the main cause for things being so hard to cross over from simulation to reality? Sensor noise? (And/or quality?) We're decent at building world models that don't permit impossible behaviour. And we're decent at training agents in those world models. What's unpredictable is how a sensor in the real world will map to a world model.

sgt101 2y ago

Simulations are closed, the real world is open. Weird things happen in the real world. They happen all the time and we expect them to happen and hardly notice when they do. Simulations approximate the model that humans use to reason about the world, but we just ignore all the oddness and can't articulate the oddness until it's made explicit by some failure.

hwillis 2y ago

Human fingertips and skin in general are extremely precise high resolution pressure and vibration sensors that robots almost completely lack. Tactile feedback takes up the second most human sensor bandwidth behind vision. Imagine doing anything without it- interacting with the world through a puppet. The joints in a human arm are direct-drive series elastic actuators with pretty incredible sensitivity. If you were holding a big heavy box and I bounced a ping pong ball off it, you'd probably feel the vibration. That's kind of a big deal. You may not be able to measure absolute weight well, or even relative weight, but you can detect quick changes with amazing precision. Lifting any heavy or large object is inherently wobbly even for a robot, and that sensitivity is important.

alexey-salmin 2y ago
This. I often feel it's not appreciated how damn brilliant a human is when you look at it as a robot. Thousands of servos, millions of sensors, high-throughput compute AND all of that self-assembled and self-repaired AND built mostly out of water, carbohydrates and like 5 grams of iron. A meatbag is by far the best robot on the face of earth today, the gap is astonishing. My estimate is that we'll get AGI faster than human-capable robots competing with meatbags cost-wise. We'll have to deploy self-assembling robots throughout the whole supply chain from sand and ore to chips and servos, eliminate human salaries everywhere, only then the price will be comparable to a parent assembling and training a child which was the way for millions of years.

0xf00ff00f 2y ago

Also relevant: https://en.wikipedia.org/wiki/Moravec%27s_paradox https://en.wikipedia.org/wiki/Moravec%27s_paradox

krisoft 2y ago

> What could be the main cause for things being so hard to cross over from simulation to reality? When you are simulating you are leaving out of details. Some knowingly, some without you even being aware of them. Let’s imagine you are working on some perception algorithm for a lidar. You have a cool simulation using geometry to calculate how far each laser goes before they bounce back. Did you remember to implement that the lidar takes some amount of time to sweep around? It is not making instantenous global captures. (Which is a lot easier to code down.) This causes a kind of smearing effect particularly as things move fast relative to the sensor. Do you know if your sensor sweeps with one plane of lasers or multiple? What is the exact geometry? Are you simulating it as the sensor is designed or as it is built? Did you remember that the sun blinds the sensor? And if so, do you simulate it correctly? Including the algorithms it has internally to try to compensate for it? Did you know that retroreflective surfaces interact with the laser differently? In particular if there is moisture on the lidar. The light straying through water droplets cause false readings. Did you know that lidars can, under certain meteorological conditions see the exhaust of diesel vehicles? I hope you planed a fluid dynamics engine into your similation to account for the exhaust plume spreading upwards and expanding. And these are just the niggles of a particular sensor modality from the top of my head. I know similarly long lists for other sensor types. And then we didn’t even got into thermal issues, or the sensor aging or anything. And I bet reality has an even longer list than I do. Is any of this important in your domain? If yes you ignore them at your peril. If no, you will waste your time implementing them in your simulation. But the only way you will know for sure is when you stumble uppon them and they surprise you out of nowhere by breaking your assumptions and your code. > Should we invest some more in stochastic programming to handle all the little quirks that cannot be modeled perfectly? That is robotics. This is what people writing software for robots do. This is what researchers working on robotics problems are working on. If you want a nice and small example of this maybe read up onnKalman filters or the RANSAC algorithm. If you fancy a more academic treatment of the topic I recommend Probabilistic Robotics from Sebastian Thrun. > is the major issue that robots must work together with humans That certainly does not make it easier. Have you seen the video of the russian chess robot breaking its oponent’s finger?[1] That is just the least gruesome of the recent incidents that I can recall. How much is this your “main problem” depens on your problem domain though. 1: https://youtu.be/Mt-1smlom_M?si=19sOPkQf3s6JHoFo https://youtu.be/Mt-1smlom_M?si=19sOPkQf3s6JHoFo

YeGoblynQueenne 2y ago

>> What could be the main cause for things being so hard to cross over from simulation to reality? Resolution. Simulations are too low-resolution and the current dominant approach to learning behaviour (Reinforcement Learning) is too high-resolution (more precisely, too high-variance). So the simulations don't represent reality accurately enough and the learned policies overfit to the simulations' inaccuracies. Learning from pixels sounds like a great idea until you realise the physical world is not quite made of pixels.