Researchers from ETH Zurich developed a robotic system that can solve a real-world labyrinth game using reinforcement learning.
As detailed in their study “Sample-Efficient Learning to Solve a Real-World Labyrinth Game Using Data-Augmented Model-Based Reinforcement Learning,” this AI-powered robot mastered the BRIO labyrinth game in just five hours of training data, outperforming any known previous attempts.
The BRIO labyrinth game, which might be familiar to some, is a test of fine motor skills and spatial reasoning that requires players to navigate a steel ball through a maze by tilting the playfield.
Despite its apparent simplicity, the game is complex due to the relationship between the ball and walls, surface irregularities, and nonlinear control knob dynamics. These challenges make the labyrinth ideal for applying and evaluating state-of-the-art robotic learning methods.
The ETH Zurich team, led by Thomas Bi and Professor Raffaello D’Andrea, developed a method that extracts efficient observations from the maze using camera images.
The AI’s learning process is based on model-based reinforcement learning, using a reward function defined by progress through the labyrinth.
After training, the AI robot successfully navigated the labyrinth with a 76% success rate and an average completion time of 15.73 seconds. This is slightly better than the best human record of 15.95 seconds.
How the study worked
The system uses a camera to capture top-down images, extracting crucial data like ball position and labyrinth layout. Machine learning techniques mirrored observations to enhance the training data, generating more diverse data and improving generalization.
This research represents a substantial step forward in applying AI in dynamic, real-world environments. The ETH team plans to open-source their project, believing that their system could serve as a valuable real-world benchmark for further AI research due to its low space requirements, modest cost, and simple hardware setup.
Further findings are published on this website, and you can watch a video of intriguing robot system works below.
One of the study’s co-authors, Professor Raffaello D’Andrea, commented, “We believe that this is the ideal testbed for research in real-world machine learning and AI. Prior to CyberRunner, only organizations with large budgets and custom-made experimental infrastructure could perform research in this area.”
“Now, for less than 200 dollars, anyone can engage in cutting-edge AI research. Furthermore, once thousands of CyberRunners are out in the real-world, it will be possible to engage in large-scale experiments, where learning happens in parallel, on a global scale. The ultimate in Citizen Science!”
Applying advanced AI systems to practically useful robotic systems has been of massive interest. Researchers recently used AI to build a robot capable of autonomously fabricating an oxygen catalyst from rock samples, and DeepMind collaborated on an autonomous research lab capable of discovering and synthesizing compounds.
ETH Zurich’s AI robot demonstrates the potential of advanced AI techniques in solving real-world challenges, bridging the gap between AI’s theoretical capabilities and its practical use in physical environments.
In the future, these technologies will come together to enable efficient, intelligent robotic systems that autonomously handle complex real-life tasks.
