The 3D view needs WebGPU or WebGL2.

inputoutputco-processor

Humanoid locomotionIn development

Walk

The fly brain learns to walk on two legs.

Walk connects the fly brain to a simulated human body with 21 joints and 42 muscles. The goal is to walk on command, stay upright when hit and get up after a fall.

Joint angles, balance and the walking command reach the ascending neurons, which report the state of the body in the fly. The descending neurons drive one pair of opposing muscles per joint.

The adapter learns from a teacher that controls the same body. The teacher learns from recordings of human movement.

21JointsMuJoCo, 2 ms physics step
42Musclesone opposing pair per joint
50 HzBrain stepsone per control step

Wiring

Brain
  • FlyWire v783, fixed weights
Inputs
  • joint angles and velocities (21 + 21)
  • balance: gravity, torso rates, head height
  • hands and feet relative to torso, foot heights
  • efference copy of 42 muscle activations
  • command: heading, speed
  • nearest logs
Outputs
  • muscles: 42 excitations, 50 Hz
Co-processors
  • none