Robotic Boomerang Throwing
via Model-Based Release Design

Yang Liu1Colin Jones2Aude Billard1

1 Learning Algorithms and Systems Lab (LASA), EPFL
2 Automatic Control Lab, EPFL

1×

Model-Based Release Design

Robot launch design

Transient release dynamics: >1,700°/s spin increase in ~100 ms

Original slide illustration of arm motion, grasp angle and in-hand boomerang spin

Predict release spin
Adjust grasp angle θg and throw motion q(t)

Shared interfaceRelease state
V0   ω0

Boomerang geometry design

Flight dynamics at low Reynolds numbers

Original slide illustration of a spinning boomerang curving through flight

Predict flight changes under design changes
Adjust size and thickness

Iterative design process for a returning throw

1×

Nominal · ~90° grasp · 45 rad/s

Need more spin to precess → Increase grasp angle

View measured trajectories
Three measured 3D flights: baseline, increased release spin, and thinner boomerang

Boomerang flight precession intuition

Fast spin generates lift → Spin lift serves as centripetal force

Source

Challenge: Dynamic mismatch with humans

Redesign boomerangs for robots

Human throw 1×

Source

Robot throw 1×

Approximate dynamic capabilities
CapabilityHumanDynaArm
Arm length~80 cm110 cm
Elbow speed~40 rad/s~6× faster7 rad/s
Wrist speed~25 rad/s14 rad/s
Hand speed~30 m/s~10 m/s
Human capabilities: rapid overhand throwing [1]

Spin-averaged flight dynamics

Rapid spin lets us average wing loads over one revolution [2]

Spin is much faster than precession and inclination change

Staged identification

Spin dynamics isolate the spin-axis torque

  1. Fit the spin-axis torque dependence
  2. Retain it while fitting all six coefficient functions

Predicting flight changes across boomerang designs

Keep the aerodynamic coefficient functions fixed
Update mass, radius and inertia

Nominal, thin and large boomerangs

Nominal · Thin · Large

0.6 mm · 0.4 mm · 0.6 mm
85 mm · 85 mm · 106.25 mm radius

Measured and predicted trajectories: the thin boomerang turns tighter, the large boomerang turns wider
Thin: tighter turn · Large: wider turn
Solid: measured · Dashed: predicted · No refitting

Transient release dynamics

Fast in-hand spinning during a ~100 ms release
The Sliding Pivot model [3] predicts the release state

Robot release 1/4×

Release model

Model screening & robot validation

Screen motion parameters in simulation → Test the predicted spin response on the robot

Vary the throw parameter
Seek a large, monotonic spin change

Predicted release spin

Simulated release spin across operating conditions, with grasp angle providing the largest spin authority
Grasp angle provides the greatest spin authority
Gripper delay gives the least consistent spin response