Google DeepMind achieves impressive results Training small off-the-shelf robots to compete in soccer matches. In a recent article in Science Robotics, the researchers detailed their innovative approach to using deep reinforcement learning (deep RL) to teach a simplified version of this movement to a bipedal robot.
Unlike previous experiments that focused on four-legged robots, DeepMind’s work demonstrates significant progress in training bipedal humanoid machines to perform dynamic physics tasks.
DeepMind’s deep reinforcement learning framework has well-documented success in mastering games like chess and Go. However, these achievements mostly involved strategic thinking rather than physical coordination. By applying deep reinforcement learning to a soccer robot, DeepMind demonstrated its ability to effectively tackle complex physical challenges.
Engineers initially trained the robot in computer simulations, focusing on two key skills: getting up from the ground and scoring goals against opponents. By combining these skills and introducing simulated game scenarios, the robot learned to play a full one-on-one soccer match. Through iterative training, they gradually improved their abilities, including kicking, shooting, defense, reaction to opponent’s movements, etc.
During testing, robots trained with deep reinforcement learning demonstrated superior agility and efficiency compared to non-adaptive scripted robots. They exhibit emergent behaviors such as spinning and spinning, which are challenging to preprogram. However, these tests relied solely on simulation-based training, and future goals are to integrate real-time reinforcement training to further enhance the robot’s adaptability.
While the technology shows promise, there are still some hurdles that need to be overcome before DeepMind-powered robots can compete in events like RoboCup. Scaling the robot and perfecting its capabilities will require a lot of experimentation and refinement. Nonetheless, DeepMind’s pioneering work highlights the potential of deep reinforcement learning to improve the locomotion and adaptability of bipedal robots in real-world scenarios.
Submitted under . Learn more about robotics and robots.