Embodied Intelligence Challenge

[nemorathwald]

Robotics advancement has accelerated vastly this decade, in Japanese research and development, and the DARPA Grand Challenge. Today I listened to an episode of the Brain Science Podcast concerning Embodied Intelligence and got thinking about how interesting it would be to put it to a real test.

If I had a hundred million dollars I might spend it on a research program not based on a specific application, but on a basic platform for applications of embodied intelligence. The goal would be a small-to-medium robot that could model its own body and the environment well enough to reliably learn how to recharge itself in arbitrary urban environments as we keep moving the power supply around. It would be increasingly difficult; at the top of stairs, behind a wall of glass, in the middle of a net stretched over a pit, and so forth. The technology platform should be adaptable to optimize locomotion for many body plans.

• The body has fine-grained sensory input.
  • stereoscopic cameras
  • stereoscopic microphones
  • infrared ladar
  • proprioception of every body segment, based on inexpensive Wiimote accelerometers
  • multi-touch-sensitive resistant and capacitive body surfaces
• The body is operated wirelessly by a computing cloud.
  • Builds candidate 3-D simulations of its body and environment based on feedback from the real world.
  • Plans actions in the modeled environments by playing them like a computer game.
  • Feels "happy" or rewarded when it achieves its goals:
  • Hunger: Recharge
  • Curiosity: Make and verify simulations of immediate future outcomes, that successfully correspond to real outcomes. (This is pretty destructive. It will probably have to take things apart to understand them. Just for one example, to find out whether a material will support its weight without breaking, it probably has to try to break a sample of it.)
  • Tests those simulated futures by operating the robot.
  • Uses feedback to eliminate incompatible candidate simulations and generate improved models of its body & environment.

More abstract forms of artificial intelligence can proceed from that foundation.

In a totally unrelated direction, I'd love to see a robot whose eyes are webcams and other senses have output displayed on a webpage. The robot could be tele-operated by the votes of thousands of people on the web who act as a sort of hive mind like the game of "Audience Pong".