What is sim-to-real?
Sim-to-real is the transfer of a policy, controller or learned behavior from a simulated environment to a physical robot. Simulation provides scale, repeatability and access to conditions that can be costly to reproduce in hardware.
Where the gap comes from
- Dynamics: real friction, compliance, backlash and actuator behavior differ from simplified models.
- Perception: cameras and sensors contain noise, latency and calibration error.
- Contacts: grasping and collisions are difficult to model perfectly.
- Environment: real objects and surfaces vary beyond the simulator.
How researchers reduce the gap
Common strategies include domain randomization, system identification, sensor modeling, robust policies, adaptation and limited real-world fine-tuning. The aim is to make behavior tolerate uncertainty rather than depend on one idealized simulation.
Why embodiment matters
Simulation is also a model of the robot. If morphology, actuator limits or sensor placement are inaccurate, a policy can fail even when the simulated task looks correct.
Simulation plus real data
A practical pipeline often combines simulation with demonstrations and deployment data. Simulation can provide breadth while real interaction provides grounding. The balance depends on the task, hardware and available data.
Read our sim-to-real research guide, then explore humanoid robot learning and robot data.