Does new correction data help improve performance?

Goal: See if new data collected from teleoperation during robot execution can be added to the DP training and improve performance.

• Collect correction data during operation
  • If necessary, ask Valentyn for help, changing controller to stop AI model when teleop occurs
  • See if we can just do adding the teleop during DP evaluation or if this results in bad trajectories
  • Record the correction data
• Re-train with the newer data
  • Training with just new data seems like the best approach
  • Evaluate the resulting policy, check performance