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.
Also known as: domain randomization, DR
Updated
What is randomised
Visual randomisation can vary textures, lighting, camera pose, object pose and clutter. The original robotics study by Tobin and colleagues trained an object localiser on rendered images whose camera positions, lighting and textures changed across samples.
For control, the distribution can also cover mass, friction, damping, actuator gains, latency, sensor noise, external pushes and initial states. Peng and colleagues used dynamics randomisation to train a pushing policy that transferred to a physical arm. Modern tooling exposes these choices directly: Isaac Lab's event functions include randomisation of rigid-body mass, materials, actuator gains, joint parameters, gravity, colour and texture.
How it addresses the reality gap
Instead of relying on one supposedly exact simulator, domain randomisation trains across a family of plausible simulators. The model is encouraged to use features or behaviours that remain effective as nuisance variables change.
This differs from system identification, which estimates parameters to make a simulator match measured hardware. The methods can be combined: identification can centre or narrow the randomisation ranges, while randomisation covers residual uncertainty.
Limits
The randomisation distribution is an engineering assumption. If it omits a real failure mode, such as backlash, deformable contact or timing jitter, the trained policy may remain brittle. Ranges that are implausibly broad can also make learning unnecessarily hard or produce conservative behaviour.
Success in randomised simulation is therefore not evidence of transfer by itself. The final test is performance and safety on representative hardware, including conditions near the intended deployment limits.
Sources
Related terms
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.
Simulation & transfer
Synthetic data
Synthetic data is training or evaluation data produced computationally rather than measured directly from the target real-world event. In robotics it can include rendered sensor observations, automatically generated labels, reconstructed or generated scenes, and simulated trajectories. A simulated rollout is one subtype: a temporally linked trajectory produced by executing a controller or policy in a simulator.
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.