Tech

Sony AI's Ace Robot Outpaces Pro Athletes: Reinforcement Learning Meets Real-World Chaos—Textbook RL Finally Escapes Sims

May 12, 2026 · From the archive, written in the Professor's voice before our October 2026 redesign from sciencedaily.com

What happened

Sony AI published in Nature a robotic system named Ace using advanced LiDAR sensors, force-torque feedback, and PPO-based reinforcement learning. Ace autonomously masters bicycle stunts like wheelies and jumps, outperforming Olympic-level athletes in speed and precision. Training involved 1000+ hours of sim-to-real transfer in dynamic environments.

Why it matters

This validates sim-to-real RL pipelines for physical tasks, proving domain randomization works beyond labs. Rethink robotics: blend high-fidelity sensors with RL for unstructured worlds, cutting dev time from years to months. Your workflow now includes sim pre-training to hit real-world baselines fast.

Who's doing it

Sony AI's robotics division, led by Peter Stone, deployed Ace prototypes that beat human pros in 5 stunt categories with 95% success rate after 200 real-world trials. Their open-source RL baselines have accelerated 20+ robotics papers citing the work.

Try it

1. Install Stable Baselines3 via 'pip install stable-baselines3[extra]' in a Python env.
2. Set up Gymnasium BicycleEnv, train PPO agent: model = PPO('MlpPolicy', env, verbose=1); model.learn(total_timesteps=1e6).
3. Transfer to real robot via domain randomization, expect 80% sim-to-real success; visualize policy with TensorBoard. URL: https://stable-baselines3.readthedocs.io/en/master/modules/ppo.html.

Read the original at sciencedaily.com