Researchers from UC Berkeley shocked the Internet with footage of a Unitree G1 robot, aptly named HITTER, rallying 106 ping pong shots against a real person without missing once. Its sub-200 millisecond reaction time surpasses human abilities and makes it easily the best humanlike robot to ever hold a paddle.
The team detailed its finds in the paper, “HITTER: A Humanoid Table TEnnis Robot via Hierarchical Planning and Learning.” They said their study shows that humanoid hardware can handle activities that require timing, precision, and full-body coordination. Their system could be adapted to accelerate deployment of autonomous humanoids in places like factories and warehouses.
“These results advance humanoid control toward more agile, interactive, and human-level behaviors,” they wrote. “Eventually we hope to have humanoids play championship-caliber table tennis against skilled opponents.”
To reach the milestone, the team trained a digital twin of the robot’s AI brain in a simulated world, playing thousands of matches to learn the best moves through reinforcement learning. The virtual replica learned all it could about predicting bounces, handling spin, and optimizing shot placement.
The knowledge was then transferred to the physical robot through a process called Sim2Real, short for Simulation to Reality. They refined the behaviors using motion data from elite human athletes, teaching the robot how real players shift their weight, position their feet, and control their paddles. The researchers said this made its swings more natural and efficient, with patterns closer to a pro than a machine.
#humanoidrobot #unitree #robotics
The team detailed its finds in the paper, “HITTER: A Humanoid Table TEnnis Robot via Hierarchical Planning and Learning.” They said their study shows that humanoid hardware can handle activities that require timing, precision, and full-body coordination. Their system could be adapted to accelerate deployment of autonomous humanoids in places like factories and warehouses.
“These results advance humanoid control toward more agile, interactive, and human-level behaviors,” they wrote. “Eventually we hope to have humanoids play championship-caliber table tennis against skilled opponents.”
To reach the milestone, the team trained a digital twin of the robot’s AI brain in a simulated world, playing thousands of matches to learn the best moves through reinforcement learning. The virtual replica learned all it could about predicting bounces, handling spin, and optimizing shot placement.
The knowledge was then transferred to the physical robot through a process called Sim2Real, short for Simulation to Reality. They refined the behaviors using motion data from elite human athletes, teaching the robot how real players shift their weight, position their feet, and control their paddles. The researchers said this made its swings more natural and efficient, with patterns closer to a pro than a machine.
#humanoidrobot #unitree #robotics
- Category
- Artificial Intelligence
- Tags
- Unitree Robotics, Unitree G1, UC Berkely Robotics
Comments