news.inema.pro / All news

community Published on 14 October 2026

Talk with Tiza about how I can help

INEMA post about measured gains in humanoid training with Real2Sim, checked against sources and added to the robotica-real2sim project.

Real2Sim is a way to train humanoid robots in virtual environments that imitate real places and tasks. In the robotica-real2sim project, the process described starts by filming a kitchen or factory, creating a digital copy of the environment, and preparing tasks for virtual training. The result is then validated before it reaches the robot. INEMA’s publication brings together measured gains and explains where simulation can help.

For users, simulation makes it possible to try walking, balancing, and fine hand movements many times without exposing a real robot to falls or dangerous situations. In one NVIDIA example with GR00T N1, synthetic data combined with real data led to 40% better performance than real data alone. This can expand training, but it does not remove the need to validate on the robot: weight, friction, and motor delays can change the result, and training for one body does not directly apply to another.

To get started, consult the humanoid section of the robotica-real2sim project. Notice how the virtual environment is created and how tasks are validated before testing on the real robot. The publication also points out important limits: simulation may not reproduce a slipping cup, an unfamiliar floor, or other physical differences accurately. The time and cost of teaching a humanoid a new task remain an open question.

Read at the source →

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