Hierarchical Training Helps a Robot Clear Tight Openings on Its Own
Robots with four legs are gaining a new kind of agility. Engineers from the University of Hong Kong and the Oxford Robotics Institute have built a system called ConsJump that lets a quadruped robot size up an obstacle, select the right motion, and jump through narrow slits without human input. The approach relies on hierarchical reinforcement learning, with two layers working together: a low-level controller and a high-level controller.
At the lower level, the robot draws on a library of agile maneuvers including walking, galloping, and jumping. These movements were trained through imitation learning and then refined with inverse kinematics. At the higher level, the controller picks the appropriate maneuver and issues velocity and turning commands. Thanks to this setup, the robot can cross obstacles close to its own body size, and it can also adjust to uneven terrain or foot slip just before takeoff.
Specs, Testing, and What Comes Next
For environmental perception, the robot relies on an Intel D435i RGB-D camera, which detects obstacles, calculates the center of an opening, and analyzes its geometry. The ConsJump system was demonstrated on Unitree Aliengo, a 22-kilogram quadruped. During the tests, the robot reached 2.5 meters per second before leaping, with a flight phase lasting about 0.44 seconds.
This advance could significantly strengthen the autonomy of quadruped robots, making them more effective in demanding settings. The research team from Hong Kong and Oxford emphasizes that such technology will play an important role in future robotics. Earlier, DaxAI Robotics showcased its four-legged robot horse Qiji X1 at the World Robotics Conference in Beijing.
ConsJump marks a notable step in robotics because it shows a robot independently adapting to difficult conditions. The implications extend to several practical uses, including:
- search operations
- rescue missions
- exploration of hard-to-reach areas
Better autonomous control can substantially improve both the effectiveness and safety of these machines in real-world scenarios. Such capabilities matter especially as robots are deployed into disaster zones and other environments where human guidance is limited.
As quadruped robots like the one developed by the University of Hong Kong gain new capabilities, it's interesting to see how similar innovations are emerging in the field. For instance, a recent presentation by a Chinese startup has unveiled a robotic horse designed to carry riders, showcasing the versatility and potential of robotic mobility. This development highlights the growing trend of integrating advanced robotics into everyday applications, further enhancing their functionality. To learn more about this innovative robotic horse, visit this article on rider-carrying robotics.