Agent Design Intuition

Input Param Intuition

• To build truely Model Free RL for the current agent we will need following information on which agent learns :
  • RGB values of the image showcasing robot from multiple positions at each timestep:
    • Need this because we want the robot to learn to corelate from visual features with its own joint positions.
      • Observation.left_shoulder_rgb: np.ndarray
      • Observation.left_shoulder_depth: np.ndarray
      • Observation.left_shoulder_mask: np.ndarray
      • Observation.right_shoulder_rgb: np.ndarray
      • Observation.right_shoulder_depth: np.ndarray
      • Observation.right_shoulder_mask: np.ndarray
  • Joint positions:
    • This is basic. Observation.joint_position()
  • Final Position Coords:
    • This is because the agent should learn to make decisions knowing whats the final goal positions. task.target.get_position()

NN Architecture Intuition

Immitation Learning NN

todo

Immitatioon Learning as base with RL

todo

Experimantation Metrics

1. Number of Steps to Goal Position post Learning.
2. Rewards in RL setup