#robotics
TLDR:

• nominal controllers - where do you get those?
• skill-chaining
• Failure Discovery -hand designed here but there are works which show that they can be learnt. They simulate the nominal policies from various initial positions to discover and record the failures.
  • impedance based controllers
• Traditional recovery learning methods are highly dataset dependent - hence they propose an online RL approach that does not rely on known preconditions.