The blind robot learned to walk for the record three hours

The blind robot learned to walk for the record three hours

In the biomemetics laboratory of the Massachusetts Institute of Technology, a speed robotic mini-hepard was introduced; researchers did not train it for specific movements, but the blind device learned to walk on its own, and then broke the speed record of such devices. It developed a speed of about 14 km/h.

Usually, in the course of robot training, they explore the real system and create a similar model by hand, but it takes about 100 days of intensive work to do it, which is the amount of time that a person needs to spend to create models and correct mistakes.

The reduction of the training to three hours required both a change in methodology and the use of NVIDIA-based simulations.

The new simulation is a safe playground where a robot can trip and fall, so the device moves faster and more intense, and learns from its mistakes. The key point of the project was simulation and the ability to make mistakes without consequences.

The MlT team has created a method to improve robot behavior through simulation experience. This has proved to be critical to the application of learned behavior models in the real world. Now, a mini-gepard can run at the same speed as a person moves.