Real-Time NMPC Wins AI Olympics' Double-Pendulum Contest

The fourth AI Olympics with RealAIGym concluded at IJCAI 2026 in Bremen on August 19-21, with organizers declaring a controller built by Nick Karydakis and Konstantinos Chatzilygeroudis, of the University of Patras' Laboratory of Automation and Robotics, the winner. Two teams reached the onsite final. The Patras entry ran a real-time nonlinear model-predictive controller (NMPC) solved by sequential quadratic programming; the runner-up, from Chalmers University of Technology in Sweden, paired its own system identification of the hardware with a separate NMPC solver tracking a precomputed swing-up trajectory.

The task is to swing up and balance an underactuated two-link pendulum from a dead stop, in either an 'Acrobot' configuration, where only the second, outer joint is driven, or a 'Pendubot' configuration, where only the first joint is driven. Unlike most AI benchmarks, it runs on physical hardware rather than in simulation: competitors submit controller code that executes remotely on the organizers' 'CloudPendulum' rig. The final evaluation spreads runs across four separate physical pendulum cells, so a controller has to cope with the small mechanical differences between machines, not just one calibrated setup. Each trial lasts 300 seconds, and the score is the total time the pendulum stays balanced inside the upright goal region, averaged over 16 runs per team — four cells times four attempts each. A separate set of runs adds an undisclosed disturbance, a push mid-swing, to test robustness.

The Patras controller unifies swing-up and stabilization into a single optimization policy rather than switching between separate controllers for each phase, using structure-exploiting ADMM and interior-point methods (HPIPM) to keep the solver fast enough for real-time hardware control — an average update rate above 100 Hz, close to 400 Hz on the Pendubot and 240 Hz on the harder Acrobot configuration. In the team's own published validation, the Pendubot controller reached 100 percent success both with and without a mid-trial push, holding the pendulum in the goal region for a median 92.5 percent of an undisturbed attempt and 87.3 percent of a disturbed one. On the Acrobot, harder because the driven joint is farther from the pivot, it held 100 percent success undisturbed and 70 percent with a push.

The Chalmers entry took a different route: a twelve-parameter friction and drag model fit to hardware data collected at 200 Hz, feeding an acados sequential-quadratic-programming NMPC solver running at a 400 Hz target rate against an offline reference trajectory. During development the team relied on a small passive joint-assist device, adding about 5 millinewton-meters of torque, which competition rules do not permit in scored runs. Without it, in the official four-cell, 300-second format, the team's own published numbers show why the approach placed second: a mean of 80 seconds of upright time out of 300 on the Pendubot and 75 seconds on the Acrobot, with the Acrobot scoring zero on two of the four cells and seven of the sixteen scored trials ending in a safety-limit stop. The team itself labeled these figures diagnostic rather than qualifying, since the assist that had made the controller work in the lab was absent on competition day.

The result is notable against the competition's own history. Earlier editions, held alongside IROS and ICRA, were dominated by controllers trained with reinforcement learning. This year's winner instead used classical trajectory optimization, run fast enough and structured well enough to compete directly on real hardware rather than a simulator. The runner-up's failure mode is the more instructive half of the story: a controller that worked reliably in the lab lost most of its margin once a development-only assist was removed for scoring, and that gap only showed up because the benchmark tests on physical rigs with real friction and real safety limits, not a simulated approximation of them.

Prizes for the 2026 edition totaled more than 2000 dollars, awarded at a ceremony on August 21. The organizers have run the competition annually since 2023, each time on the same double-pendulum hardware, which keeps year-to-year comparisons meaningful even as the entrants' methods change.