Quadruped robot developer Deep Robotics has introduced a new series of AI-enabled robots featuring advanced reinforcement learning techniques. These robots showcase enhanced dexterity, autonomy, and self-balancing capabilities.
The company presented its cutting-edge technology at the IEEE International Conference on Robotics and Automation (ICRA), held last week in Yokohama, Japan.
One of the highlighted models, the Lite3, demonstrated impressive abilities by running across a stack of boxes, jumping over gaps, and navigating stairs. The robot also exhibited a cat-like ability to readjust its center of gravity when thrown into the air, landing on its feet.
To achieve these capabilities, Lite3 was trained using multiple AI simulations. Its avatar was initially programmed to always be on a slant, enhancing its center of gravity control. "We successfully applied the Proximal Policy Optimization (PPO) method combined with depth camera sensor inputs and trained an end-to-end policy," explained Zunwang Ma, a robot control engineer at Deep Robotics. This method allows the robot to traverse high platforms and gaps without requiring precise perception and robot modeling.
The larger-scale robot, the X30, was also on display, showcasing its autonomous capabilities on various terrains such as grass, gravel, and rubble. Designed to withstand external interference, like pushing or pulling, the X30’s balance maintenance is a significant advantage for outdoor operations. Potential applications for these robots include autonomous inspection, emergency detection, and rescue missions.
Deep Robotics plans to incorporate advanced AI technologies into its robotic systems under its "AI-Plus Plan," aimed at supporting innovations in embodied AI. The company intends to expand its product capabilities to include perception, planning, decision-making, and human-machine interaction. Ma emphasized the growing role of reinforcement learning in the company’s robots to enhance terrain adaptation and robustness. "With reinforcement learning, we can use fewer policies to achieve greater terrain adaptation and coverage, making the robot more robust and intelligent," Ma stated.
