Project · Accepted at IEEE-RAS Humanoids 2026

Tight Spiral

A humanoid robot that throws a tight-spiral American football.

  • Unitree G1 + Dex3-1
  • OptiTrack motion capture
  • Model-based control and optimization

Joint work with Bowen Weng.

01

What a tight spiral is

At release, the ball's spin axis, its nose, and its velocity must line up. I score each throw with two metrics.

Spin efficiency
100% means all of the spin is about the ball's long axis.
Nose angle
0° means the nose points exactly along the velocity.

02

Why it's hard

  • Coupled momentum, fast. The robot has to reach the required coupled momentum with high-DoF motion in about half a second.
  • A brief, underactuated release. The release lasts under 100 ms while the ball rolls partly across the fingers, leaving it effectively underactuated.
  • Joints at their limits. In the final 50 ms of one throw, waist pitch was torque-saturated 90.9% of the time, and shoulder pitch and elbow were saturated 54.5% of the time.

03

How it works

  1. 1

    Legs

    A whole-body controller (GR00T) keeps the robot balanced. An extended Kalman filter estimates pelvis motion from contacts, leg kinematics, and the IMU.

  2. 2

    Throw phase

    Trajectory optimization (Crocoddyl + Pinocchio) runs offline on a 10-DoF waist-and-arm model over 0.6 s, and the robot tracks the result. Target launch elevation: 35°.

  3. 3

    Follow-through

    Model predictive control drives the wrist and the fingers still in contact, with a 15-step horizon of 4 ms steps. One online solve took about 500 ms, so I run it as an offline lookup table. It penalizes off-axis spin, nose misalignment, and fingers pushing into the ball.

  4. 4

    Release

    The thumb opens at the end of the throw while the other fingers stay in contact. Follow-through lasted 78.5 ms on average over 61 trials.

04

Results

  • 93.6%
    Peak spin efficiency
  • 16.4°
    Best nose-angle error0.286 rad
  • 5.35 m/s
    Peak release speed
  • 14.5 rad/s
    Peak spin rate

Ablation Without follow-through, average spin efficiency fell from 67.1% to 60.9% and average nose angle rose from 18.9° to 34.7°.

05

Setup

Robot
Unitree G1, 29 DoF
Hand
Dex3-1, 7 DoF, three fingers, with a 3D-printed extended thumbnail
Ball
Mini football: 0.252 kg, 21.6 cm long, 12.7 cm wide, 12 psi
Tracking
OptiTrack, six PrimeX 22 cameras at 360 Hz, ≤0.15 mm accuracy

06

Cite

Z. Mahboob and B. Weng, "Throwing a Tight Spiral American Football by a Humanoid Robot," accepted at IEEE-RAS Humanoids 2026. arXiv:2608.16642.