Solving Dexterous Manipulation Tasks with Trajectory Optimisation and Reinforcement Learning

Henry Charlesworth, Rowden Technologies

Bristol ML Meetup - 25/06/21


  • Most robots in industry use parallel jaw grippers to manipulate objects.
  • Developing autonomous robots that can perform a wider variety of tasks in unstructured/uncertain environments will require more sophisticated manipulators.
  • The human hand is probably the most versatile and sophisticated manipulator we know of.
  • Natural to try and create robotic hands based on a human hand - and to try and train those robotic hands to perform complex manipulation tasks!

Why this is Hard?

  • Complex, discontinuous contact patterns between the hand and object make it difficult to come up with an accurate Physics model to optimise.
  • High-dimensional inputs/actions, and generally high precision and coordination between many joints is required to perform tasks well.
  • Traditional robotic control approaches struggle with these kinds of problem.

Reinforcement Learning

Gradient-free Trajectory Optimisation

Made with