Collision Randomisations for Sim2Real
Project description
The main approach for transferring a deep reinforcement learning policy learnt in simulation to a real platform is domain randomisation. This is a proven method however it is found to be computationally and time intensive.
Another recent approach has looked at adding perturbational forces to elements of the system and found it to be more robust when transferring. Another similar approach that extends these and looks to satisfy the need for higher fidelity collision resolution in simulation is the randomisation of the collision meshes of robots in simulation.
By altering the collision meshes throughout training the policy will learn to overcome similar disturbances in the real system.
We look to explore this further by training a visuomotor manipulator policy to stack cubes in simulation whilst randomising the gripper contacts. The policy will then be transferred to a real robot. This will build on top of a code-base readily available.
More information is available here and on the GitHub project page.
Requirements and eligibility
We expect to see a record of a satisfactory academic standard.
Required skills: Programming (C++, Python) and experience with robot simulators.
How to apply
Industrial internships are not part of a formal program so to apply please contact Dr David Howard via david.howard@csiro.au to discuss your interest.
Please supply the following information and documentation at time of contact for your application to be considered:
- a current c.v.
- details of grades or an academic transcript
- the degree you are enrolled in
- number of years completed