news.inema.pro / All news

project Published on 14 October 2026

robotica-real2sim: 50 episodes become a LeRobot dataset in 69 s

robotica-real2sim turns 50 episodes into a LeRobot dataset in 69 seconds.

robotica-real2sim is an open project that recreates an AI robotics workflow on a computer. From a photo of a table, it builds a physics simulation in MuJoCo with a Franka Panda arm. An A4 sheet in the photo helps estimate the objects’ size and position. The project does not use Gemini 4 Argon; it recreates the idea with open tools.

The simulation can generate training data without a person guiding the robotic arm through every attempt. In the project’s recorded result, 50 varied episodes were completed in 8.7 seconds. Each episode is annotated with steps and events, as well as a quality label. Before a plan is sent to a real robot, the system tests it 20 times in the simulation; plans with an accuracy below 90% are rejected. This makes it possible to experiment without owning a robot, but the results depend on how well the simulation represents the physical world.

To get started, the guide calls for Python 3.10 or later, Git, and MuJoCo; Codex is optional for interpreting the photo and writing the plan. Take the photo from above and include an A4 sheet as a scale reference. The project also lets you create the scene manually. Check that the table’s digital twin matches the photo: the project itself reports contact, grasp, and force errors that could produce misleading results. Simulation can help assess a plan, but it needs to be checked.

Read at the source →

The link opens the original content; some require a subscription.