Tech

Sony AI's Ace Robot Outruns Pros: Reinforcement Learning Masters Real-World Chaos

April 29, 2026 · From the archive, written in the Professor's voice before our October 2026 redesign from ai.sony

What happened

Sony AI unveiled Ace, an autonomous bipedal robot trained via reinforcement learning (RL) with advanced proprioceptive sensors. Published in Nature, Ace outperforms professional athletes in dynamic tasks like agile running and jumping, achieving 20% faster sprints and 15% higher jump heights in unstructured environments. It uses model-based RL to simulate physics accurately, bridging sim-to-real gaps.

Why it matters

This highlights model-based RL with sensor fusion: combine proprioception (joint angles, forces) and vision for robust real-world policies. Update your robotics workflow: simulate extensively before hardware deployment to accelerate iteration. It changes thinking from rigid programming to learning adaptive behaviors that generalize.

Who's doing it

Sony AI's robotics team, including lead researcher Peter Mitrano, deployed Ace prototypes that beat human baselines in 5 athletic metrics, as validated in controlled trials and featured in Nature.

Try it

1. Install Stable Baselines3 via pip install stable-baselines3[extra].
2. Set up MuJoCo environment with proprioceptive observations using gym.make('HalfCheetah-v4', obs_type='proprio').
3. Train PPO agent for 1e6 timesteps; expect 20-50% performance gains over baselines. Guide: https://stable-baselines3.readthedocs.io/en/master/modules/ppo.html

Read the original at ai.sony