Imitating
a project I saw on the web (by Tyler Streeter at Iowa State), I built a system for training a humanoid robot to balance. It didn't learn very well, but it looked cool. It definitely produced some interesting behaviors. It generated some very convincing seizures and death scenes. Tyler's was better.
He typically fell over a lot. But he tried to counter balance.
I attempted to hard-code a solution that would let the humanoid balance. My solution didn't work so well. It tended to overreact... a lot. When it started to tip one direction, it would violently flail in the opposite direction; usually that meant that he flipped and landed on his head.
One of the best solutions the learning approach could find was to gracefully fall into a sitting position. Then it was still somewhat upright, but not unstable like an inverted pendulum. He looked like he was tired of trying. Amanda always felt so bad for the character whenever she came over. The poor thing spent a lot of time writhing on the ground and Amanda's empathy couldn't bear it.
He typically fell over a lot. But he tried to counter balance.
I attempted to hard-code a solution that would let the humanoid balance. My solution didn't work so well. It tended to overreact... a lot. When it started to tip one direction, it would violently flail in the opposite direction; usually that meant that he flipped and landed on his head.
One of the best solutions the learning approach could find was to gracefully fall into a sitting position. Then it was still somewhat upright, but not unstable like an inverted pendulum. He looked like he was tired of trying. Amanda always felt so bad for the character whenever she came over. The poor thing spent a lot of time writhing on the ground and Amanda's empathy couldn't bear it.
No comments:
Post a Comment