Artificial intelligence
Sim-to-real transfer
Definition
Sim-to-real transfer applies a model, policy, or behavior developed in simulation to a physical system. Its central challenge is the difference between the simulated environment and the robot, sensors, and interactions encountered in reality.
Also known as: Sim2real, Simulation-to-reality transfer
Updated
Simulation and hardware differ
A simulated robot can have different friction, actuation, sensing, or contact behavior from its physical counterpart. Peng and colleagues describe how policies can exploit simulator-specific dynamics and fail when those assumptions change on hardware.
A pushing policy, for example, may learn exactly how far a simulated object slides. The same command can produce a different result with a heavier object or another surface.
Transfer can target perception or control
Tobin and colleagues train visual object localization using randomized rendered images and demonstrate its use in real robot grasping. Peng and colleagues instead randomize simulated dynamics while learning a robot control policy.
These are different transfer problems: one concerns interpreting images, while the other concerns choosing actions under changing physical behavior.
Training strategies need hardware evaluation
Domain randomization varies simulated conditions so the learned system experiences a broader range. System identification can instead help estimate a model from physical measurements; adaptation methods can use target-domain data.
A simulated success rate is not a real-robot success rate. Report the physical tasks actually tested and any calibration or additional training used during transfer. Successful transfer in one pushing setup does not establish transfer for every contact-rich humanoid skill.
Sources
Related terms
Domain randomization
Domain randomization varies properties of training environments to encourage a learned model or policy to work across changing conditions. In robotics it often randomizes simulated appearance, physical parameters, or both to support transfer to real hardware.
Domain adaptation
Domain adaptation adjusts a learned model to work on a target data distribution that differs from its source training distribution. Robotics examples include adapting perception from simulation to camera images or adapting behavior to changed physical conditions.
System identification
System identification estimates a model of a physical system from measured inputs and outputs. In robotics, it can recover parameters such as inertia and friction or learn a more general model of how actions change the system state.