DEEP Robotics Reports Navigation Gains With New Embodied AI Models

HANGZHOU, China -- DEEP Robotics has announced five research projects into embodied AI, claiming success in developing robot navigation and control thanks to the creation of a single AI architecture for robots that function in dynamic environments.

These projects were disclosed by the company on Oct. 24 and include world models, vision-language-action systems, reinforcement learning, predictive spatial memory, and robot evaluation datasets.

One of the best results was achieved in DR-APO project, which is focused on reinforcement learning and robot control improvement at different AI architecture.

DEEP Robotics claims that the method helped increase the robot navigation success rate on LYNX M20 wheel-legged robot from 25 percent to 65 percent in real-world trials. In the simulation, the success rate grew from 70 percent to almost 90 percent.

Neither these results have been independently reproduced nor did the company prove that the improvements will be sustainable in the commercial operation conditions.

Another project – DR-WM-1.0 – focuses on helping robots predict the physical impact of their movement. The system is different from other visual prediction generators, as it takes into account such features as friction, roughness, and stiffness of the robot's interaction with the environment.

The company claims that there were employed 2.58 million image-mask pairs and over 100,000 real-robot video clips in the research.

Also, DEEP Robotics presented DR-VLA-2.0. The system allows training the navigation model with the use of human video and robot demonstration. The manipulation framework employs language commands and force feedback for robotic arm movement adaptation during physical interaction.

The other projects relate to spatial memory and evaluation. DR-Agent predicts object location changes. On the other hand, DR-Benchmark consists of LynxScenes dataset acquired through the LYNX M20.

The projects were conducted by the researchers from such universities as MIT, Tsinghua University, and Zhejiang University. At the moment, there is no information on when the full system is expected to be released or if the research advancements were implemented in the robots used by customers.