Sim-to-Real Transfer
A technique for training a robot or agent in simulation and then deploying the learned policy in the real world, bridging the reality gap.

Sim-to-real transfer (often shortened to sim2real) covers the set of techniques for training an agent — typically a robot — in a simulated environment and then deploying the learned policy in the real world without costly retraining. The motivation is simple: simulation is a thousand times faster and infinitely less risky than the physical world. The challenge is just as simple: no simulation is perfectly faithful, and a policy trained on idealized physics often fails when it meets real friction, sensor noise or changes in lighting. The gap between the two worlds is called the reality gap, and all the sophistication of sim-to-real lies in bridging it.
The reality gap in practice
A physics simulator (MuJoCo, Isaac Sim, Habitat) approximates reality with models: numerical integrators for dynamics, point contacts instead of real contact areas, Coulomb friction, sensors with no noise or latency. These approximations are necessary to reach simulation rates in the range of 1,000-100,000 Hz, but they introduce systematic discrepancies with the real world.
A policy trained to exploit these approximations learns behaviors that don't exist in reality — for example, pushing an object by relying on contact dynamics specific to the simulator. On transfer, the policy fails. This is known as the sim2real gap.
Three main families of techniques
Domain randomization
The most widely used technique. At each training episode, the physical parameters are randomized: object mass (±20%), friction coefficients, motor latency, sensor noise, colors and textures. The policy never sees exactly the same physics twice, so it can't memorize a specific configuration — it has to learn behaviors that are robust to noise. If the randomization covers a wide enough cloud, the real world looks like one of the possible simulations, and the policy generalizes.
OpenAI pushed this idea with Automatic Domain Randomization (ADR) in its Rubik's Cube project (2019): the algorithm gradually widens the randomization ranges as the agent masters simpler versions — a self-generated curriculum. In the end, the equivalent of 13,000 years of simulated experience was needed, but the result held up on a real robotic hand, even when a human prodded the object with a stuffed giraffe.
Domain adaptation
Instead of making the policy robust to noise, you learn a transformation between simulated and real observations. For example, a GAN learns to turn simulator images into "real-looking" images or, conversely, to project real images into the simulated distribution the policy knows. A more complex approach than randomization, and rarer in practice.
System identification + fine-tuning
You measure the real robot's physical parameters (masses, friction, motor latency) through targeted trials, calibrate the simulator accordingly, train on it, then run a short fine-tuning on a few hours of real episodes. This is the dominant approach in 2025 for industrial deployments: a pragmatic trade-off between robustness and precision.
The 2025 simulation ecosystem
Four simulators shape research and industry.
MuJoCo (DeepMind, open source since 2021) remains the academic reference for articulated dynamics and contact. Its GPU version MuJoCo-Warp (2025), developed jointly with NVIDIA, speeds up workloads by a factor of 70x — making it possible to train robotics foundation models (GR00T N1) on tens of millions of episodes in a few days.
NVIDIA Isaac Sim / Isaac Lab combines PhysX physics simulation, photorealistic RTX rendering and massive parallelization (up to 16,000 simultaneous environments on a single H100). It's the de facto industrial standard for humanoids (Tesla, Figure, 1X and Apptronik all use Isaac).
Habitat (Meta AI) is optimized for navigation and household interaction with photorealistic 3D scans (Replica, HM3D). It serves as the base for Spot AI (Boston Dynamics + Meta) and much embodied AI research.
CARLA and Waymax dominate autonomous driving — Waymax (Google, 2024) adds simulation of 1,000 simultaneous vehicles to train planning policies.
Limitations and current debates
Sim-to-real remains imperfect on three fronts.
Rich contacts and deformable objects — simulators model sticky friction, fabrics, cables and liquids poorly. Bimanual manipulation of soft materials (folding laundry, untying a knot) remains a challenge for 2025-2026.
Proprietary sensors — a Hesai lidar, a DIGIT tactile sensor or an event camera are not finely modeled by standard simulators. Policies trained on generic models transfer poorly.
Real data is still required — the community is converging on a hybrid approach: 90% simulation for pre-training, 10% teleoperated real data for final fine-tuning. Sim-only remains rare in production.
Standards and Specifications
MuJoCo (Multi-Joint dynamics with Contact)
Open-source physics simulator maintained by Google DeepMind, the de facto standard for robotics research. Used by OpenAI (Rubik's Cube), DeepMind (RoboCat) and most academic labs
NVIDIA Isaac Sim / Isaac Lab
GPU-accelerated simulation platform for robotics, integrated with Omniverse. Supports photorealistic rendering (RTX), PhysX physics simulation and massive parallelization (thousands of simultaneous environments)
Habitat (Meta AI)
Photorealistic 3D simulator for navigation and interaction in home environments — the training base for Spot AI (Meta + Boston Dynamics) and much embodied AI work
Frequently Asked Questions
Why train in simulation rather than directly on the real robot?
Three reasons. Speed: a GPU-accelerated simulator can generate 100-1,000 times more episodes per hour than a physical robot, and run in parallel across thousands of simultaneous environments. Cost: a $50,000 humanoid that breaks during a rough trial is expensive; in simulation, a crash is free. Safety: you can test exploratory policies (RL) without putting a human or equipment at risk.
The flip side is the reality gap: simulation is always imperfect (approximated friction, idealized sensors, simplified contact dynamics), so a policy that is perfect in sim can fail in reality. Sim-to-real transfer techniques exist precisely to bridge that gap.
What is domain randomization?
Domain randomization means training the policy while randomly varying the simulator's physical parameters at each episode: object mass, friction, colors, initial positions, sensor latency, motor noise. The agent never sees exactly the same physics twice.
The bet: if the policy works across a wide cloud of different simulations, the real world will look like one of those possible simulations, and the policy will generalize. It's the approach that allowed OpenAI to solve the Rubik's Cube with a robotic hand (2019). For that project, OpenAI developed Automatic Domain Randomization (ADR), which gradually widens the randomization ranges as the agent masters the simpler versions — a self-generated curriculum.
Is the reality gap still a problem in 2025?
Yes, but its nature has changed. Photorealistic simulators (Isaac Sim, Habitat 3.0) and accurate physics models (MuJoCo-Warp, Isaac Lab) have sharply reduced the gap for perception and rigid-body dynamics. The areas that remain hard: contact-rich tasks (manipulating deformable objects, fabric, cables, liquids), proprietary sensors (lidar, specific tactile sensors) and environmental surprises (a person walking by, lighting that changes abruptly).
The dominant 2025 approach is hybrid: massive pre-training in simulation (Isaac Lab, MuJoCo-Warp) + short fine-tuning on real data (1,000-10,000 teleoperated episodes) — far more effective than sim-only or real-only.
What are the major simulators?
MuJoCo (DeepMind, open source) — the academic reference for articulated dynamics and contact. Used by OpenAI, DeepMind and Stanford. Isaac Sim / Isaac Lab (NVIDIA) — massively parallel GPU-accelerated simulation, photorealistic RTX rendering, the top industrial choice. Habitat (Meta) — a photorealistic 3D simulator optimized for navigation and household interaction. SAPIEN (UC San Diego), RoboCasa (NVIDIA), ManiSkill (Hillbot) — manipulation environments.
For autonomous driving: CARLA (Intel Labs, open), Waymax (Google), NVIDIA DRIVE Sim. Each simulator has its biases — a robot moved from one simulator to another isn't guaranteed to work without readjustment.