By learning by strengthening, researchers teach a humanoid robot to avoid objects launched at high speed without losing balance or adopting unrealistic postures.
Humanoids gain in agility, but to work alongside humans they must also know how to react to the unexpected, as an object that arrives towards them. A team of seven researchers presents Humanoid Dodgeball, a reinforcement learning framework that drives a humanoid to avoid rapid projectiles through body-wide movements, while remaining stable.
The difficulty lies in the balance between three often contradictory demands: react quickly to the threat, keep natural movements and remain robust despite the dynamic complexity of a body with many joints. The authors have designed a reward that takes into account the trajectory of the threat, and several opposing "a priori of movement", one per flight direction, learned from human movements, which direct the robot towards plausible and varied gestures rather than effective but unrealistic contortions. The simulation tests then on a real robot show a clear progress in avoidance for different shooting speeds and distances.
Read the full article on the original source
Go to the source →