Questions: Sim-to-Real Transfer and Domain Adaptation
1 questions to test your understanding
Score: 0 / 1
Question 1 Multiple Choice
A robot learns dexterous hand manipulation in PyBullet (physics simulator) with realistic friction, contact dynamics, and object properties. After training for 100k episodes, the policy achieves 95% success at picking up objects in simulation. Deployed on real hardware with the same hand and objects, success drops to 30%. Which factor is most likely responsible for the largest performance gap?
AThe neural network is too large and overfits to simulation
BSimulation physics are idealized and deterministic; real friction is variable, contacts are unstable, and actuators have latency and backlash not modeled in simulation
CThe real hardware has broken sensors
DThe learning algorithm (RL) is inappropriate for real robots
Domain randomization is a major success story in applied robotics. Google, OpenAI, and others have published results where policies trained on heavily randomized simulation transfer to real hardware with high success rates on dexterous manipulation, navigation, and other tasks. The insight — that robustness to parameter variation is more important than parametric accuracy — has been validated empirically many times and has become standard practice.