3 ms·
I wonder if Deepmind's quest for AGI using games as testbed will translate in the real world. It is quite possible that many of those feats rely on the ANNs lea
by seydor 3y ago
I wonder if Deepmind's quest for AGI using games as testbed will translate in the real world. It is quite possible that many of those feats rely on the ANNs learning the physics engine of the game. Which as impressive as it is, it 's probably nowhere near as complicated as the real world out there. The last mile may prove to be a very hard problem to solve, and i wonder if they have an alternative strategy to it.
- bigyikes 3y agoAre real world physics really that much more complicated than video game physics? Sure, if you’re talking about E&M, relativity, or QFT, the real world is more complicated, but for every day mechanics, they seem pretty comparable? Video games might be even harder to predict sometimes, since the physics simulations have very strange edge cases. There are no physics glitches in the real world.
- seydor 3y agoi believe it is as hard as it is for robots to act exactly as they do during simulation
- Jensson 3y ago> Are real world physics really that much more complicated than video game physics? Massively, in a game running into a wall is perfectly normal and valid strategy to get close to it, in reality that will wreck you. Or running on a fist sized rock has no consequences, in reality that destroys your foot. Reality is full of such extreme threats everywhere even in normal homes.
- sailfast 3y agoWouldn't a general purpose 3d AI learn just as quickly about those issues in the real world as they do in-game issues? Video game physics can also be made more difficult - it's not a stretch to think about a reality simulator that dials these real-world effects up to 11 for AI to train in, at which point you need the meatspace bot and many things start to happen.
- tjah1087 3y agoSIMA author here - SIMA is betting on simulations and games. We use real-world open-ended language, but games/sim obviously have simplified physics/graphics. In terms of alternative strategies, Google DeepMind also has an amazing robotics team with lots of fantastic work for real-world robotics - including multi-robot generalists, showing positive effects when co-training one agent or model on multiple environments/bodies. Their prior work was very inspirational to us in SIMA! https://deepmind.google/discover/blog/scaling-up-learning-across-many-different-robot-types/ https://deepmind.google/discover/blog/scaling-up-learning-ac...
- hnhn34 3y agoSorry if this is a stupid question but even after reading OP’s post, I’m having trouble understanding the limitations of this. What would happen if, for example, the system were trained on simple 3D puzzle games, with natural language instructions such as “solve this puzzle” with a hint like “X or Y strategy might work”? I see that menu navigation is part of the training set. Can this thing learn to “read” or does it just learn the results of menu actions? If it can learn to play No Man’s Sky.. can it learn to play Zelda?
- robrenaud 3y agoSimulation to reality is a huge bottleneck in a lot of reinforcement learning work. Reality is just super messy and complicated. Tesla has an alternative. If you can get your devices widespread and can be recording observations and actions, you can collect huge datasets in the real world. The first embodied, vaguely general, multi-task, useful and economical robots might just really open up the virtuous cycle of experience, learning, feedback and improvement. If I had to guess where it would come from right now, I'd pick Amazon warehouses.