What is sim-to-real transfer in robotics?
Sim-to-real transfer means training a robot's control software in a computer simulation and then running it on a physical robot. The difficulty is that simulated and real conditions differ, a mismatch known as the sim-to-real gap.
Also known as: Sim2Real, Simulation-to-real transfer
Researched and fact-checked by AI, with no human review. 8 sources listed below. How we verify
Last updated
How it works
Engineers build a virtual copy of a robot and its surroundings. A learning algorithm practices there by trial and error. The resulting control software, called a policy, is then loaded onto the real machine. A 2020 IEEE symposium survey gives the reasons: real-robot data is costly to gather, while a simulator can supply as much as needed with fewer safety concerns. It also names the drawback: the gap between simulated and real worlds makes a policy perform worse on a physical robot.
Domain randomization is one remedy the survey lists. In a 2017 paper, Josh Tobin and colleagues varied how the simulator rendered its images. That way, real camera views would look to the model like one more variation. They reported that a detector trained only on simulated images located real objects to within 1.5 centimeters.
Why it matters for humanoid robots
Training on real legged robots is complicated and expensive, according to a 2019 Science Robotics study. Its authors trained a policy in simulation for the four-legged ANYmal robot. They reported the robot ran faster than before and recovered from falls. In a 2023 paper, Ilija Radosavovic and colleagues reported training a humanoid walking controller in randomized simulated environments. They reported deploying it "zero-shot," meaning with no further real-world training, to walk over varied outdoor terrain.
OpenAI said in October 2019 that neural networks trained entirely in simulation let a robot hand solve a Rubik's Cube. It put the success rate at 60% on scrambles needing 15 rotations and 20% on the hardest scrambles. Because friction, elasticity and dynamics were hard to model, the company said, it developed a technique that keeps producing harder simulated environments.
Where things stand in 2026
Nvidia's June 2026 announcement of a reference humanoid design for academic research lists its Isaac Sim and Isaac Lab software for simulating, training and testing policies before real-world deployment. In April 2026, IEEE Spectrum relayed a claim from the robot maker Agility that sim-to-real training lets its team teach the Digit humanoid new whole-body skills overnight.
Developers also describe limits. Scott Kuindersma, Boston Dynamics' vice president of robotics research, told IEEE Spectrum in September 2025 that adding simulation data to training had improved performance on the real robot. But he said models still need enough high-quality on-robot data. He said no one knows the right ratio of real to simulated data.
Sources
- Sim-to-Real Transfer in Deep Reinforcement Learning for Robotics: a Survey, arXiv (Zhao, Queralta and Westerlund; 2020 IEEE Symposium Series on Computational Intelligence)
- Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World, arXiv (Tobin et al., submitted to IROS 2017)
- Learning agile and dynamic motor skills for legged robots, arXiv (Hwangbo et al., published in Science Robotics, 2019)
- Real-World Humanoid Locomotion with Reinforcement Learning, arXiv (Radosavovic et al.)
- Solving Rubik’s Cube with a robot hand, OpenAI
- NVIDIA Announces NVIDIA Isaac GR00T Reference Humanoid Robot for Academic Research, NVIDIA
- Video Friday: Digit Learns to Dance—Virtually Overnight, IEEE Spectrum
- Large Behavior Models Are Helping Atlas Get to Work, IEEE Spectrum
