Domain Randomization for Sim2Real

Key Concepts

• Randomize visual, dynamics, and sensor parameters.
• Use curriculum to expand ranges gradually.

Common Pitfalls

• Randomization too wide makes policy underfit.
• Mismatch between randomized assets and target robot.

Practical Checklist

• Start with camera and lighting randomization.
• Add dynamics later (mass, friction, delay).
• Track real-world success rate per setting.

Math Note

Reward shaping often uses: \(r_t = \alpha r_{task} + \beta r_{stability}\)