invisible fΦrce Knowledge hub
Learning

What is the
sim-to-real gap?

The sim-to-real gap is the difference between how a robot behaves in simulation and how it behaves in the real world. A policy that works perfectly in the simulator can stumble, overheat or stall on the real machine, because the simulator left something out.

Where the gap comes from

A 2025 survey of sim-to-real methods organises the problem by which part of the robot's decision loop the simulator gets wrong: what the robot observes (rendering and sensor noise), how the world responds to its actions (contact, friction, motor dynamics), and how actions are executed (delays, torque limits).[1]

Some gaps are about mechanics: real feet slip, real cables sag. Others are about physics the simulator does not model at all. Motors heat up and lose torque; batteries sag under load; a magnetic gripper's holding force depends on the steel it touches. A simulator that never computes temperature cannot teach a policy to avoid overheating.

A concrete, measured example

A May 2026 paper trained a Unitree A1 quadruped to walk while carrying a 3 kg load. With a standard policy that ignored temperature, the robot's motors overheated after about 5 minutes. When the authors added a whole-body thermal model to training, the robot walked for more than 13 minutes.[4] This is the kind of physics Invisible Force adds to simulators.

How the gap is closed

Domain randomization. During training, the simulator randomly varies things it is unsure about: masses, friction, motor strength, even gravity and image textures. The policy learns to cope with the whole range, so the real value is just one more case. The idea dates from 2017, when it was applied to images only.[6] Today NVIDIA Isaac Lab ships randomisation functions for mass, centre of mass, friction, actuator gains, joint properties, gravity and random pushes.[5]

Learning the correction. ASAP, a 2025 paper from Carnegie Mellon and NVIDIA, records a real Unitree G1 humanoid, measures how its motion differs from the simulator's, and trains a small "delta" model that corrects the simulator. Agile whole-body skills then transfer to the real robot.[2]

A live twin. Real-is-Sim (2025) keeps a simulator running in sync with the real robot at all times, a dynamic digital twin, so that a single policy sees the same thing in both worlds.[3] See Digital twin vs. simulation.

Better physics. The most direct fix: put the missing physics into the simulator, so there is less to randomise and less to correct. That is Invisible Force's approach.

References

  1. Da et al., Arizona State University, arXiv 2502.13187. A Survey of Sim-to-Real Methods in RL: Progress, Prospects and Challenges with Foundation Models. February 2025.
  2. He et al., Carnegie Mellon University and NVIDIA, arXiv 2502.01143. ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills. February 2025.
  3. Abou-Chakra et al., arXiv 2504.03597. Real-is-Sim: Bridging the Sim-to-Real Gap with a Dynamic Digital Twin. April 2025.
  4. Wan et al., arXiv 2605.27046. Learning to Balance Motor Thermal Safety and Quadrupedal Locomotion Performance with Residual Policy. May 2026.
  5. NVIDIA. Isaac Lab documentation: isaaclab.envs.mdp (randomisation functions). 2026.
  6. Tobin et al., OpenAI and UC Berkeley, arXiv 1703.06907. Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World (origin). March 2017.

Every link was opened and checked in September 2026. Links open in a new tab.