XYZ Embodied AI, a company based in the Beijing Economic-Technological Development Area (BDA), also known as Beijing E-Town, has worked with universities and research institutions to develop two AI models designed to improve robots' ability to navigate and understand their surroundings. The research, known as DA-Nav and AdaOcc, addresses two key challenges in deploying robots in open, dynamic environments: making reliable navigation decisions and building accurate three-dimensional representations of the space around them.
DA-Nav focuses on translating human navigation instructions into spatial decisions that robots can execute. Conventional robot navigation often relies on simultaneous localization and mapping (SLAM), in which a robot first builds a map, determines its position, and then plans a route. In large outdoor environments, however, this approach can be costly because of the data collection required. Maps can also quickly become outdated as conditions change, with parked vehicles and moving pedestrians altering the surroundings.
DA-Nav takes a different approach by treating continuous navigation planning as a visual target-selection task. It divides the forward-facing camera image into a virtual grid of 29 columns and 24 rows, from which the model selects a sequence of target points to guide the robot.
The model also uses a chain-of-thought reasoning mechanism to structure its decisions. It first assesses whether it has deviated from the intended route, then determines the appropriate action and finally generates specific trajectory points. This sequence of assessment, decision-making and execution is designed to help the robot correct its course when deviations occur.
In tests, the research team deployed the trained model directly on quadruped and humanoid robots without additional fine-tuning on real-world data. The robots completed more than 1,200 meters of stable navigation across different urban environments, demonstrating the model's ability to transfer to real-world settings without prior adaptation to those environments.
AdaOcc addresses a complementary challenge: three-dimensional perception. The model can adapt to different sensor configurations and computing budgets, allowing it to operate across different types of robotic platforms. Tests have demonstrated that a single pretrained model can support multiple tasks in real-world environments, including building obstacle maps for navigation and locating target objects for manipulation tasks. It can also be used with different platforms, including quadruped robots, humanoid robots and robotic arms.
Together, DA-Nav and AdaOcc address two practical requirements for embodied AI: translating visual and semantic information into spatial decisions, and enabling robots to perceive and navigate environments despite variations in sensors, hardware and real-world conditions. By accounting for these constraints in the model design, the research offers an approach to making robotic systems more adaptable beyond controlled laboratory settings and better suited to practical applications.