Visualizations of the controlling of the learned robots of SARD
Point Navigation
The agent is generated at the center of a flat arena and needs to reach a random goal (red square) in this arena. This type of symmetric body allows it to move in any direction.
Escape Bowl
Generated at the center of a bowl-shaped terrain surrounded by small hills, the agent has to escape from the hilly region. This type of symmetric body makes it more stable and less likely to fall down.
Patrol
The agent is required to run forth and back between two target locations along the x-axis. This type of symmetric body enables it to run forth and back without having to turn around.
Locomotion on Variable Terrain
The goal of the agent is to maximize forward displacement on a variable terrain over an episode. This type of symmetric body makes it more stable.
Locomotion on Flat Terrain
Similar to locomotion on variable terrain tasks, the agent is initialized on flat terrain. This frog-like symmetric body has never appeared in previous works, and it runs faster.
Manipulate Box
The agent is required to move a box (small cube) from the initial position to the target place (red square). This frog-like symmetric body has more potential to finish this challenging task.
Visualization of the learning process of SARD in Patrol Task
T=5M
The robot still had trouble in running to the goal ahead.
T=25M
The robot learned to reach the goal ahead but fails to learn to run backward.
T=30M
The robot learned to reach the goal ahead and behind without turning around due to its symmetry.
T=50M
The robot learned to run faster between two goals without the need to turn around at the end of the training, which indicates the effectiveness of symmetry.