Tech

Sony AI's Ace Robot Outpaces Pro Athletes: Reinforcement Learning Triumph

May 8, 2026 · From the archive, written in the Professor's voice before our October 2026 redesign from ai.sony

What happened

Sony AI published in Nature a breakthrough with Ace, an autonomous bipedal robot using advanced force/torque sensors and model-based reinforcement learning. Ace beats professional athletes in agile tasks like high jumps (1.5m) and 400m sprints (faster than human elites). It handles dynamic real-world environments with zero-shot generalization. Source: https://ai.sony/news/sony-ai-announces-breakthrough-research-in-real-world-artificial-intelligence-and-robotics.

Why it matters

This demonstrates sim-to-real transfer via reinforcement learning with domain randomization, bridging simulation gaps for physical AI. You now prioritize sensor fusion and RL over pure vision models for robotics. Workflow changes: simulate extensively before hardware deployment to accelerate iteration.

Who's doing it

DeepMind's RoboCat, led by researchers like Raia Hadsell, self-improves on 100+ tasks via RL fine-tuning, achieving 3-5x faster adaptation than baselines on real robots.

Try it

1. Install Isaac Gym via NVIDIA's GitHub: git clone https://github.com/isaac-sim/IsaacGymEnvs.
2. Train a bipedal walker with PPO algorithm, adding domain randomization (e.g., friction 0.5-1.5).
3. Deploy to hardware via sim-to-real scripts; expect policy success rate >90% in dynamic environments. URL: https://github.com/NVIDIA-Omniverse/IsaacGymEnvs.

Read the original at ai.sony