Simulation & transfer
Sim-to-real
Sim-to-real is the transfer of a model, policy or behaviour developed wholly or partly in simulation to a physical robot or real environment. The central problem is the reality gap: errors in simulated dynamics, sensing, appearance and timing can make a strategy successful in simulation but unreliable or unsafe on hardware.
Also known as: sim2real, simulation-to-real transfer
Updated
What is transferred
The transferred artefact may be a perception model trained on rendered images, a control policy learned through simulated interaction, a planner tested against virtual scenes or parameters initialised from simulation. Deployment then runs that artefact against real sensor observations and real robot dynamics.
Sim-to-real describes this direction of transfer, not a particular algorithm and not a guarantee of zero real-world data. Some systems deploy directly from simulation; others calibrate, validate or fine-tune with measurements from hardware.
The reality gap
Simulation necessarily simplifies the target system. Contact models, friction, compliance, actuator delay, camera response, sensor noise and scene appearance can all differ from reality. Peng and colleagues describe how control behaviours can become specific to simulator dynamics and fail after deployment because of modelling error.
The gap is especially consequential for humanoids. Small errors in contact timing, mass distribution or actuator response can affect balance and whole-body motion, while visual mismatch can change what a learned policy perceives.
Common transfer methods
System identification fits simulated parameters to measured behaviour. Domain randomisation trains across parameter variation; the original visual domain-randomisation study varied rendering so real images could fall within the learned distribution. Other approaches use realistic rendering, representation or domain adaptation, real-data fine-tuning, residual control and teacher-student distillation.
These methods address different mismatches and are often combined. Evaluation should compare the same task in simulation and on representative hardware, report any real data used during adaptation, and test disturbances not seen during training. A high simulated return alone says little about safe physical deployment.
Sources
Related terms
Simulation & transfer
Domain randomisation
Domain randomisation is a simulation-training technique that varies selected visual, physical or sensor parameters across generated examples or episodes. By exposing a model or policy to a distribution of environments rather than one calibrated scene, it aims to make real conditions fall within the learned variation and improve sim-to-real transfer.
Simulation & transfer
Real-to-sim
Real-to-sim is the process of using observations or measurements from a real robot, object or environment to construct or update a simulation-ready representation. Depending on the application, this may recover geometry, articulation, calibration and physical parameters, or translate real demonstrations into simulator state, usually to support later analysis, training or sim-to-real deployment.
Models & learning
Policy
A policy is the decision rule that maps a robot’s current observations or estimated state, and sometimes a task instruction, to an action or probability distribution over actions. It can be hand-designed or learned from demonstrations, rewards or both. In humanoid robotics, its outputs may be joint targets, torques, end-effector changes or higher-level skills.
Models & learning
Reinforcement learning
Reinforcement learning is a method in which an agent learns a policy by interacting with an environment and optimising cumulative reward. In humanoid robotics, actions change the robot and world, while observations, rewards and episode endings provide experience for improving balance, locomotion or manipulation behaviour.