A Multi-stage Approach for Efficiently Learning Humanoid Robot Stand-up Behavior

Dingsheng Luo

Abstract:

Stand-up motion is among the most essential behaviors
for humanoid robots. For achieving stable stand-up behavior,
the traditional key-frame based motion planning methods are
time-exhausted and expert knowledge dependent. On the other
hand, classic trial-and-error based learning methods are inefficient
due to the high degrees of freedom (DOFs) for humanoid
robots and the difficulty in fixing appropriate reward functions. In
this paper, a multi-stage learning approach is proposed to address
the above issues. At the first stage, under a trajectory based
motion control model, key motion frames sampled from human
motion capture data (HMCD) are used for model initialization,
through which the solution space could be pruned. At the second
stage, the design of experiments (DOE) technique is introduced
for fast and active searching in the pruned solution space. At
the last stage, a refining process that adopts a stochastic gradient
learning strategy is performed to achieve the final behavior. Under
this three-stage learning framework, along with a simple heuristic
reward function, the learning of the stand-up behavior for a kidsize
humanoid robot is fulfilled successfully and efficiently.