Sim2real Gap

Transfer learning

Considerations for Simulation

1. System identification
• Replicating the input-output behavior or a dynamical system
• Finding exact physics parameters is long-standing problem

2. Domain randomization
• environment randomization is one of solution
• source environment
• target environment
• physics randomization with reasonable range is one of solution
• visual randomization (visual transformation)

3. Domain adaptation

Reasons why Sim2real Gap happens

• Exact physical parameters like friction, gain, density, deformable objects are unknown.

Insights for future works

• The only ultimate way to resolve this issue is that we need real-time training for new environment like human does.