RAI's AthenaZero throws a baseball at 113 km/h
The pitch is a stunt. The number under it is 3.97 kg of effective mass at the wrist.
AthenaZero, a robot from the RAI Institute, throws a baseball at 113 km/h. Lukas Ziegler wrote up the system, which is on this month's cover of Science Robotics.
Throwing a baseball is just the most legible way to prove you have it.
Effective mass at the wrist: 3.97 kg. The paper measures a Franka at 29.21 kg and a UR5e at 34.72 kg on the same terms.
The headline number is not the pitch speed. It is effective mass at the wrist, which is how heavy the arm feels to anything it hits. AthenaZero measures 3.97 kg. The paper puts a Franka at 29.21 kg and a UR5e at 34.72 kg on the same terms. That is close to an order of magnitude lighter, and it lands in the range of a human arm.
Where the lightness comes from
A conventional industrial arm runs gear ratios of 80 to 1 or higher, and a gearbox multiplies reflected inertia by the square of its ratio. You get a small motor and an arm that feels heavy to move. AthenaZero uses quasi-direct drive at roughly 5 to 1 on most joints, and remotises the transmissions so the heavy parts sit proximal, near the torso, instead of out at the wrist.
The rest of the build, per Ziegler. One torso joint with a single degree of freedom, two 7-DoF arms, two 6-DoF under-actuated hands. That is 22 actuators across 27 joints. It stands 1.6 m tall with a 1.8 m wingspan and a payload above 3 kg.
What it does with it
Throwing above 30 m/s. Catching at over 14 m/s across 7.3 m. Batting with an 82% contact rate at 31 mph. Ball sports are a convenient test because they force the arm to reverse direction fast and eat contact it did not fully plan for.
That is the actual argument. Low effective mass is what lets an arm change direction at human cadence, absorb a contact instead of transmitting it, and work next to a person without hurting them. Ziegler's framing is that every contact-rich task sits downstream of that property, and a baseball is simply the most readable way to show you have it.
The effective mass analysis tooling is open under an MIT licence, with datasets on Zenodo, so the comparison against the Franka and the UR5e is something other labs can run rather than take on trust.
What the sources cover is the hardware and the sports demos. They do not describe factory or home deployments, and no third party has published its own measurement of the arm yet.