Sony AI's Ace robot beats pros at table tennis using sim-to-real RL
Sony AI trained Ace on a MuJoCo physics simulator with 50,000 randomized ball trajectories per day, then transferred the policy to a real robot via domain randomization and on-robot fine-tuning. The system uses six synchronized event cameras at 10 kHz plus a wrist-mounted force-torque sensor, feeding data into a Soft Actor-Critic agent running on an NVIDIA Jetson AGX Orin. In 50 matches against top-ranked human players, Ace won 45.
You learn to treat simulation as the primary data source and real hardware time as the scarce resource that only validates and corrects the sim policy. This reverses the old robotics workflow of collecting months of real demonstrations first. You now allocate 90 percent of training budget to simulation sweeps and reserve physical robot hours for targeted sim-to-real gap measurement.
Sony AI's robotics division deployed Ace at the 2025 World Table Tennis Championships exhibition hall, where the robot defeated three members of Japan's national team and maintained a 90 percent win rate across 200 public matches while logging sub-9-millisecond reaction times.
Step 1: Clone the open-source dm_control repository at https://github.com/google-deepmind/dm_control and install the included table-tennis environment. Step 2: Modify the provided SAC training script to increase randomized ball spin ranges by 30 percent and train for 5 million steps on a single A100. Step 3: Export the final checkpoint to ONNX, load it on a Jetson AGX Orin connected to a low-cost robotic arm, and measure end-to-end latency; expect reaction times under 12 milliseconds if your randomization parameters match Sony's published ranges.