Short answer: the sim-to-real gap is the distance between the world the simulator modelled and the world the robot actually meets. Friction is never quite what the physics engine assumed, sensors are noisier, actuators respond later, floors and parts vary, and components wear. A policy trained to perfection in the model has often learned the model's quirks rather than the task. Teams close the gap by randomising simulation conditions, identifying the real system's parameters, and finishing training with real-world data — and buyers close it by demanding evidence from real floors, not demo reels from virtual ones.
The five sources of the gap
- Contact and friction. Grasping, pushing and walking are contact problems, and contact is where physics engines approximate hardest. A grip that is stable in simulation slides in reality because the friction coefficients were polite fictions.
- Sensor reality. Simulated cameras and depth sensors return clean data. Real ones suffer glare, dust, motion blur, calibration drift and dropped frames — usually at the worst moment.
- Latency and dynamics. Actuators have lag, gearboxes have backlash, structures flex. A controller tuned against instant, rigid responses can oscillate or overshoot on hardware.
- Variation. The simulated part is the nominal part. Real parts arrive with tolerances, burrs, oil films and packaging differences the model never met.
- Wear and drift. The robot that was calibrated in week one is a slightly different machine by month three. Simulation is timeless; floors are not.
How teams close it
The standard toolkit has three layers. Domain randomisation trains the policy across deliberately varied simulations — friction, lighting, masses, delays and textures all randomised — so the policy learns the task's invariants instead of one world's details; the technique was popularised in the late 2010s and is now standard practice. System identification goes the other way: measure the real robot and floor carefully, then tune the simulator to match them, narrowing the gap from the model side. Real-world fine-tuning accepts that simulation gets the policy most of the way and budgets real robot hours — supervised, on the actual task — for the last stretch, which is also where teleoperation-collected data earns its keep. Our companion piece on robot simulation and digital twins covers the training infrastructure these techniques run on.
Validation: the ladder between simulation and floor
Because the gap cannot be eliminated in advance, serious teams stage the crossing. First the policy runs in simulation across randomised conditions; then against real control hardware in hardware-in-the-loop rigs, where latency and actuator behaviour are genuine even if the world is not; then on a single real robot, supervised, on the actual task; and only then across a fleet or a production line. Each stage has exit criteria — success rates, intervention counts, worst-case behaviour — and the discipline is refusing to skip a stage because the demo looked clean. When a supplier describes their process, the presence of this ladder is itself evidence.
What this means for buyers
Sim-to-real has quietly become a procurement topic, because more of what suppliers sell — picking skills, navigation behaviour, humanoid demonstrations — is now trained substantially in simulation. The questions that expose the gap are simple ones. What was trained in simulation and what on hardware? On how many real units, for how many hours, in environments like yours? What does performance look like on day ninety, after wear and recalibration, rather than at the acceptance demo? Suppliers with real deployments answer with numbers from floors; suppliers with only a simulator answer with videos. The directory's embodied AI and robot software category (73 companies as of 10 November 2026) is where the software side of these claims can be compared company by company.
None of this argues against simulation — it is the only economically sane way to give robots millions of practice repetitions. It argues for treating simulation results as a promise that must be redeemed in the real world, and for pricing the redemption into the project from the start.
Frequently asked questions
What is the sim-to-real gap in robotics?
It is the performance drop when a robot policy trained in simulation meets the physical world. Simulators approximate contact and friction, return cleaner sensor data than real cameras, ignore actuator lag and backlash, and model nominal parts rather than real tolerances and wear. A policy can master the model while learning behaviours the real floor punishes — closing that distance is the sim-to-real problem.
What is domain randomisation?
A training technique in which simulation conditions — friction, lighting, object masses, sensor delays, textures — are deliberately varied at random, so the learned policy must handle a family of worlds rather than one exact model. If the real world falls inside the randomised family, the policy transfers. The technique was popularised in the late 2010s and is now standard practice in robot learning.
How do teams close the sim-to-real gap?
Three methods combine. Domain randomisation makes policies robust to variation. System identification measures the real robot and environment and tunes the simulator to match. Real-world fine-tuning budgets supervised hours on actual hardware for the final stretch of performance. Serious programmes use all three, plus validation testing on real floors before deployment.
What should buyers ask suppliers about simulation-trained robots?
Ask what was trained in simulation versus on hardware, how many real units and hours back the performance claims, and what results look like after months of wear rather than at the demo. Simulation-trained behaviour is legitimate and increasingly normal — the buyer's job is to price the transfer: evidence from real floors is the currency, videos from virtual ones are not.