Sony AI's Ace Robot Outperforms Pro Athletes in Real-World Tasks via Reinforcement Learning
Sony AI published in Nature the Ace system, an autonomous bipedal robot using advanced force-torque sensors and model-based reinforcement learning. Ace beats professional athletes in dynamic tasks like ball kicking and object catching. It achieves 90% success rates in unstructured environments after 100 hours of simulation training.
This validates sim-to-real transfer in reinforcement learning, bridging simulation gaps with domain randomization. Shift your thinking: real-world robotics demands sensor fusion over pure vision. Integrate this into workflows by starting with MuJoCo simulations before hardware deployment.
Sony AI's robotics division deployed Ace prototypes that outperformed human pros in 5 athletic benchmarks, including 2.5x faster ball interception speeds. Their Nature paper details zero-shot generalization to new environments.
Step 1: Install Stable Baselines3 and MuJoCo via pip install stable-baselines3[extra] (https://stable-baselines3.readthedocs.io/). Step 2: Train PPO agent on HumanoidEnv for locomotion; run 1e6 timesteps for stable walking gait. Step 3: Transfer to real robot with domain randomization; expect 80% sim-to-real success after tuning sensors like IMU data.