Slashdot Mirror
Elon Musk's Open Source OpenAI: We're Working On a Robot For Your Household Chores (zdnet.com)
An anonymous reader writes from a report via ZDNet: OpenAI, the artificial-intelligence non-profit backed by Elon Musk, Amazon Web Services, and others, is working on creating a physical robot that performs household chores. In a blog post Monday, OpenAI leaders said they don't want to manufacture the robot itself, but "enable a physical robot [...] to perform basic housework." The company says it is "inspired" by DeepMind's work in the deep learning and reinforcement learning field of AI, as displayed by its AlphaGo victory over human Go masters. OpenAI says it wants to "train an agent capable enough to solve any game," noting that significant advances in AI will be required in order for that to happen. In May, the company released a public beta of a new Open Source gym for computer programmers working on AI. They also have plans to build an agent that can understand natural language and seek clarification when following instructions to complete a task. OpenAI plans to build new algorithms that can advance this field. Finally, OpenAI wants to measure its progress across games, robotics, and language-based tasks, which is where OpenAI's Gym Beta will come into play.
Asimo takes an awkward, brute-force-it approach involving lots of complex rules and experimentally-determined data tables to control what position to move each servo/actuator to at each stage of a step relative to what feedback it's getting from its sensors. Every time you change anything about the scenario - the texture of something, the weight of something, etc - you need to go back to modifying tables and potentially adding whole new rules.
That is, of course, not how lifeforms learn to do things, and it's a serious hindrance to task diversity and development time. We don't work by saying, "bend this joint to 63.02 degrees, this one to 11.17 degrees and this one to 32.88 degrees because I've calculated that this will position your hand where it needs to be". We work kinematically - when there's a task we've never done before, we set our body in motion, get a rough estimate of where our body will end up, and increase or decrease forces, constantly re-evaluating how things our going. The more we do a task, the more our "neural net" gets used to what sort of forces will be needed to accomplish a given task and the less it needs to constantly re-evaluate and adjust. But the key is, we don't work by fixed positions and angles - we work through forces and velocities.
If we want robots to be able to interact with their world the way we do, we need to give them the sort of paradigm that we use. Give them adjustable-tensioned cable "muscles", or at least a good emulation of them. Don't precalculate what angles everything needs to end up in. Do give them the most advanced neural net you can. Do give the neural net the visual or other sensor hardware needed to get an accurate sense of where its extremities are, how they're moving, and where other objects in the scene are and how they're moving (relative to itself). With these things, you'll end up with a task-flexible robot with natural movement. Precalculated angles and positions will always have comparatively poor task flexibility and unnatural movement.
Monkeywrench Ex Machina.
This is actually going to be the biggest problem with robots that actually work in the way that sci-fi envisioned. If people can train them to do illegal acts, there's going to be a heck of a lot of government regulation and only the government and only the specially licensed and vetted who will be allowed to train them.
What chores will Rosie do? Cooking? No, way too complex for decades yet: gathering foods from fridge and pantry, opening a variety of containers, exracting indredients of many shapes and consistencies in proper amounts (and avoiding those that are spoiled), prepping each (peeling, chopping, grating, sauteeing), and synchronized cooking before dishing up. All this without burning down the house or spilling and then having to clean up the messes. Not to mention the cleaning up of utensils and pans thereafter.
Cleaning? Not even. Vacuuming, dusting, tidying, navigating a dynamic ever changing floorplan and tabletops, without toppling and destroying all those expensive knick-knacks or running over the cat, and of course, not making an ever greater mess.
Washing the dog? Nursing the baby? Cleaning the windows? Mopping the floors? Beating the rugs? Washing the car? Mowing the yard? Clearing the gutters? Weeding the garden? Trimming the hedge? Edging the sidewalk? Taking the dog for a walk? Taking out the garbage? Yeahhhh...
Which subset of these chores does OpenAI choose for their Rosie Jetson? 'Cause she can't do them all. And how many Rosies can you sell if she costs more than a Segway but only can load the dishwasher and fetch beer from the fridge?
Reality check: an affordable domestic robot that's actually useful is hellaciously ambitious, especially when no robot on earth can do any of these things yet, not at any price.