Sony AI's Ace Robot Outperforms Pro Athletes via Reinforcement Learning
Sony AI introduced Ace, an autonomous robotic system that beats professional table tennis players using advanced LiDAR sensors and model-based reinforcement learning. Published in Nature, Ace achieves rally durations of over 100 strokes and wins 80% of matches against humans. The system employs MuJoCo physics simulation for rapid policy training before real-world transfer.
This breakthrough shows how simulation-to-real transfer accelerates robotics development. Shift your thinking to hybrid model-based RL for handling dynamic environments. It enables faster iteration in your AI projects by bridging sim and reality gaps.
Sony AI's robotics team, led by Peter Kormushev, deployed Ace to defeat a top-100 world-ranked player in extended rallies, achieving superhuman spin control and speed adaptation as detailed in their Nature paper.
Step 1: Install Sony AI's open-source Ace framework from https://github.com/SonyAI-Ace/ace-robotics. Step 2: Train a policy in MuJoCo simulator using DreamerV3 RL algorithm on table tennis tasks for 1 million steps. Step 3: Transfer to a real robotic arm with LiDAR; expect 70%+ win rate against amateur players after 10 hours of fine-tuning.