IJCAI 2026 AI Olympics with RealAIGym
DFKI – German Research Center for Artificial Intelligence
📌About the Event
IJCAI 2026 AI Olympics with RealAIGym is an international AI and robotics competition focused on testing “athletic intelligence” in real-world robot control. Teams will build controllers for a double pendulum robot and compete on the task of swinging it up and stabilizing it in an upright position using real hardware via CloudPendulum. Unlike earlier editions, the 2026 challenge trains policies on real hardware from the start, making it highly relevant for students and researchers in AI, robotics, reinforcement learning, and optimal control.
ℹ️ Event Details
- Edition: 4th AI Olympics with RealAIGym
- Conference: IJCAI-ECAI 2026, Bremen, Germany
- Mode of conduct: Hybrid-style competition
- Remote qualification: Teams access the real double pendulum system online through CloudPendulum.
- On-site final: Final evaluation will happen at IJCAI 2026 in Bremen, Germany.
- Main task: Swing up and stabilize an underactuated 2-link robotic system, either Acrobot or Pendubot, in the upright position.
- Technical focus: Robotics, reinforcement learning, model-based control, optimal control, and real-hardware benchmarking.
📅 Important Dates
- Registration deadline: 31 May 2026
- Qualification stage: 1 June – 15 July 2026
- Solution/report submission: At the end of the qualification stage, including a 2–4 page report
- Qualification winners announcement: 31 July 2026
- Revised report submission: 7 August 2026
- Final competition at IJCAI 2026: 15–21 August 2026
🎯 Eligibility & Rules
- Eligibility: Primarily for students and researchers in AI, machine learning, reinforcement learning, optimal control, robotics, and related fields.
- Team requirement: Participants are expected to develop or improve algorithms from model-based control, reinforcement learning, or hybrid approaches.
- Hardware access: Teams will train on real hardware remotely through CloudPendulum, with limited access time and no model parameters provided in advance.
- Allowed feedback: Participants will receive only sensor feedback such as position, velocity, and torque from both actuators.
- Attempt limit: Each swing-up attempt must be completed within 60 seconds.
- Controller rule: The controller must inherit from the provided AbstractController class.
- Hardware limits:
- Control loop frequency: Maximum 500 Hz
- Torque limit: 0.15 Nm
- Velocity limit: 40 rad/s
- Position limits: ±360 degrees for both joints
- Restriction: Friction compensation is not allowed, increasing the difficulty of the challenge.
🏆 Prize Details and Value
- Recognition: Best-performing teams will receive prizes and special recognition.
- Conference opportunity: Top teams may be invited to the IJCAI 2026 final competition in Bremen.
- Showcase opportunity: Selected teams can showcase their trained controllers and discuss their solutions during the IJCAI competition program.
- Research value: Participants get exposure to a rare real-hardware AI robotics benchmark, useful for research portfolios in robotics, RL, and control systems.