Shadow Hand learns how to juggle with our improved CEM (iCEM) method

Shadow Hand learns how to juggle with our improved CEM (iCEM) method

In our newest paper Sample-efficient Cross-Entropy Method for Real-time Planning we are able to solve a number of challenging robotic tasks such as dexterous in-hand manipulation or the humanoid stand-up task almost in real-time while achieving state-of-the-art performance.

Our key contributions are:

  • Adding colored noise and correlations
  • Adding memory to CEM by keeping an elite-subset over multiple inner CEM-iterations and shifting an elite-subset to the next outer CEM-iteration
  • Executing always the best action instead of the mean actions, clipping action boundaries and decaying the population size

This work is a first step towards using zero-order optimization strategies for real-time planning and control.