Humanoid robots pull off full game of long rope skipping without any help
Two humanoid robots turning a rope in perfect rhythm while a third jumps through it sounds like a playground scene. Researchers at Nanjing University just built the artificial intelligence to make it happen. The team says it marks the first time multiple humanoid robots have learned to cooperate on a shared physical task this precise.
The system, called Marope , teaches robots to coordinate long rope skipping. That game normally requires two people turning a shared rope while a third jumps through it on rhythm. The researchers detailed their approach in a preprint paper and published demo videos on a dedicated project page.
Why this task is hard
Long rope skipping poses a different challenge than most robotic sports research to date. Prior humanoid athletic projects, including robots that play table tennis, dance, or run, generally involve a single robot acting alone or reacting to an external object.
Rope skipping demands genuine multi-robot teamwork instead. Two robots must swing a shared rope in a coherent motion while keeping their balance and footing. The rope itself bends and flexes unpredictably, and its shape cannot be directly controlled the way a rigid object can. On top of that, the turners must time their swings to match a jumper whose rhythm may vary.
A three-layer approach
Marope tackles the problem in stages. First, the two rope-turning robots learn a shared, decentralized policy for swinging the rope through multi-agent reinforcement learning. This lets each robot control its own movements while producing coordinated rope motion.
A separate scheduling policy then sits above that foundation, adjusting the turners' commands in real time. This keeps the rope's rotation lined up with the jumper's timing and avoids collisions. Finally, the researchers trained the jumping robot itself on a range of jumping styles, so the system generalizes to different partners rather than just one predictable jumping pattern.
Tested in reality
The researchers evaluated Marope against several simplified baseline approaches in simulation. These included a single-agent version that controlled both turning robots through one combined policy. Marope's approach reduced rope tracking errors and cut feet slippage compared to those baselines, according to the results reported in the paper.
The team also deployed the system on real Unitree G1 humanoid robots . In these tests, the setup successfully coordinated humanoid-to-humanoid rope turning and humanoid-to-human turning. It also handled jumping scenarios involving both human and humanoid participants, according to the paper and accompanying demo footage.
Limits the team acknowledges
The researchers were direct about where Marope currently falls short. The framework is built specifically for long rope skipping and does not generalize automatically to other cooperative humanoid tasks. The current system also only handles a single jumper at a time, and trickier formats like double dutch remain unsolved.
The team relies on a motion-capture system to track the rope and players in real time. That setup would need to be replaced with onboard sensors before humanoids could realistically coordinate with people in everyday settings. The researchers describe that shift toward onboard perception as a promising direction for future work.
