Prototype robot learns to jump through tight gates using vision

A 22 kg quadruped runs at a gate roughly the size of its own body. Its onboard camera decides when to commit. In mid-air, the Unitree Aliengo folds its legs tightly enough for its feet to slip past the frame, then lands running. Researchers at the University of Hong Kong and the Oxford Robotics Institute have demonstrated the dog-inspired maneuver on real hardware rather than in simulation.

The trick was not to script every joint. The team’s low-level controller learned pacing, cantering, steering and jumping from motion-capture data recorded from a real dog. A discriminator network judged whether the robot’s movement resembled the animal’s, and that judgment became part of the training reward. Walking, running, jumping and landing could then blend continuously instead of living as isolated skills.

A second controller handled the immediate decision. Running 10 times per second, it used an RGB-D camera — a camera that captures both color and depth — to detect the gate and output only forward speed and turning rate. That narrowed decision space allowed training in about six hours, according to senior author Peng Lu. The system itself discovered the take-off sequence and the mid-air leg tuck.

The behavior also adapted beyond one rehearsed position. The researchers report that the gate could be placed anywhere in the room and at any height within the robot’s reach. For a higher opening, the robot used a longer run-up and folded its joints more aggressively; when the gate was off to one side, it crossed diagonally. It could also re-plan its heading if the gate moved while the robot was already running.

And so what, concretely? A robot that can commit to a dynamic maneuver could have more routes through cluttered, human-scale environments such as collapsed structures and industrial sites, where slowly negotiating every obstacle may make a path impossible. The reusable skill library could also be retargeted to hurdles, vertical fences or other gaps by changing the high-level module rather than retraining locomotion from scratch. For now, this is a prototype demonstration on the Aliengo, and the researchers are planning further studies to improve the approach.