Real-world observation. Robot-world execution.

Real2Gym

Real-World Videos to
Executable Robot Simulations

From a human gesture or a robot demonstration to a world that robots can act in.
Reconstruct in Blender. Execute in MuJoCo. Refine with GPT6 Astra.

HUMAN & ROBOT INPUTS · BLENDER + MUJOCO · SCENE AUGMENTATION
HUMAN → ROBOT / REAL RECORDING

A mug. A microwave. A robot.

Open video ↗

Place the mug inside, withdraw the hand, and close the door. A human demonstration retargeted to dual FR3 + Franka Hand.

Native task completed
About this run 374 frames · 30 fps

01 / DEMONSTRATIONS

Different demonstrations.
One path to simulation.

Six source scenes. Eight robot embodiments and demonstrations.
Frame-matched Real RGB / Blender RGB / MuJoCo RGB.

REAL SIMULATION

Showing all 8 demonstrations.

02 / SCENE AUGMENTATION

One change.
A visibly different world.

Geometry, position, orientation, table height, textures, and distractors.
Plus three background settings. Compare each one with the original.

OBJECT GEOMETRY

Microwave scaled to 75%

The microwave is 25% smaller in all three dimensions. The action is adapted to its smaller cavity.

Native task completed
GRASP MUG PLACE INSIDE
WITHDRAW CLOSE DOOR

One category changes in each example. Both views are MuJoCo, with the same camera and event index; playback follows source timing (3× native speed). Augmented scenes have no paired real recording.

Choose a single change6 individual changes + 3 background settings
BUILD WITH REAL2GYM

Bring your own
real-world demonstration.

Explore the workflow, try a scene, or build on the framework.

Explore the code
Citation
@software{real2gym,
  author = {{Real2Gym Contributors}},
  title = {{Real2Gym: Real-World Videos to
            Executable Robot Simulations}},
  year = {2026},
  url = {https://github.com/cskrren/Real2Gym}
}