Accelerating the practical application of embodied intelligence, the Institute for AI Industry Research (AIR) of Tsinghua University launches the Embodied Simulation Platform | Frontline

Text | Wang Fangyu

Editor | Su Jianxun

To intelligently perceive the three-dimensional space like a human is one of the core capabilities of embodied intelligence. How to develop this ability and solve the bottleneck of data shortage has always been an important issue that the embodied intelligence industry wants to break through.

On December 2, the DISCOVER Laboratory of the Institute for AI Industry Research (AIR) at Tsinghua University held an Open Day event and presented the problem-solving approach of Tsinghua AIR.

At this event, the DISCOVER Laboratory launched the DISCOVERSE Embodied Simulation Platform. This platform aims to achieve universal data simulation for embodied tasks by digitizing the high-fidelity physical world, combining the physical engine and parallelized rendering technology, and for the first time, it supports rich robot tasks in different-scale indoor and outdoor scenes.

Academician Zhang Yaqin delivering a speech at the AIR Open Day event

It is introduced that the DISCOVER Laboratory of the Institute for AI Industry Research (AIR) at Tsinghua University was established in 2020, focusing on research in the field of embodied intelligence. The laboratory has been committed to exploring the Real2Sim2Real (RSR) architecture to achieve a closed-loop circulation of real-world and simulation training data.

Real2Sim2Real means from reality to simulation and then back to reality, that is, applying the simulation method to the reality and comparing their performances, and updating the simulation model and method according to the differences between the simulation and the reality. This approach can expand one piece of real data into thousands of pieces of data, thereby helping to solve the data scarcity problem faced by embodied intelligence.

In this regard, Academician Zhang Yaqin said in his speech: "In the process of moving towards the Digital 3.0 era, intelligent robots and autonomous driving have become key industrial fields. However, both are facing challenges such as high R & D costs and slow productization processes. Through the new paradigm of Real2Sim2Real (RSR), it is expected to build a bridge connecting the atomic world and the bit world, helping the industry reduce R & D costs and accelerate the practical application of intelligent systems."

The DISCOVERSE Embodied Simulation Platform is precisely built based on the Real2Sim2Real paradigm. Zhou Guyue, the director of the DISCOVER Lab, said that the DISCOVER Lab has constructed the high-fidelity simulator DISCOVERSE, successfully connecting the rendering engine and the physical simulation, and supporting end-to-end and parallelizable realistic simulation.

In addition, the DISCOVER Laboratory also launched the RDK UniBots Embodied Intelligence Science and Education Platform at the event. This platform, as the first official release of the general architecture of the embodied intelligence science and education robot.

It is introduced that the RDK UniBots platform is oriented towards the development of robots from 0 to 1, is compatible with a wide range of educational occasions, and has been applied in university science and education practices and many international robot competitions.