Learning to Throw-Flip

TL;DR

This work tackles the challenge of throwing objects to a precise landing pose by coupling impulse-momentum-based control with a data-driven learning loop. By decoupling displacement and rotation through a transient hinge release, the authors expand the feasible landing poses using a three-parameter flip family indexed by $\gamma$, $s$, and $D$, and they combine a physics-driven free-flight model with learning to adapt commands. The study demonstrates that a projectile-dynamics–aware forward model (Model 2) achieves up to 40% lower error than an end-to-end approach (Model 1) and requires fewer iterations to reach target poses, with transfer learning under CoM shifts further reducing sample needs by about 70%. Experimental results on a Franka Panda show the framework can throw-flip to a pose target within $\pm 5\text{ cm}$ and $\pm 45^{\circ}$ in dozens of trials, highlighting practical gains for speed and robustness in logistics-inspired manipulation. The approach also reuses prior in-hand spinning data to accelerate learning for new objects, offering a path toward real-world, data-efficient dynamic manipulation.

Abstract

Dynamic manipulation, such as robot tossing or throwing objects, has recently gained attention as a novel paradigm to speed up logistic operations. However, the focus has predominantly been on the object's landing location, irrespective of its final orientation. In this work, we present a method enabling a robot to accurately "throw-flip" objects to a desired landing pose (position and orientation). Conventionally, objects thrown by revolute robots suffer from parasitic rotation, resulting in highly restricted and uncontrollable landing poses. Our approach is based on two key design choices: first, leveraging the impulse-momentum principle, we design a family of throwing motions that effectively decouple the parasitic rotation, significantly expanding the feasible set of landing poses. Second, we combine a physics-based model of free flight with regression-based learning methods to account for unmodeled effects. Real robot experiments demonstrate that our framework can learn to throw-flip objects to a pose target within ($\pm$5 cm, $\pm$45 degrees) threshold in dozens of trials. Thanks to data assimilation, incorporating projectile dynamics reduces sample complexity by an average of 40% when throw-flipping to unseen poses compared to end-to-end learning methods. Additionally, we show that past knowledge on in-hand object spinning can be effectively reused, accelerating learning by 70% when throwing a new object with a Center of Mass (CoM) shift. A video summarizing the proposed method and the hardware experiments is available at https://youtu.be/txYc9b1oflU.

Figures (10)

• Figure 1: Robot throw-flips the bar with 4 different landing poses.
• Figure 2: Major notations for flipping.
• Figure 3: Schematic of flip motion. The robot firmly grasps the bar and accelerates it to a high-energy, high-velocity state, then enters a rapid decelerating-braking phase. During braking, the robot's gripper begins to open with decreasing normal force. Approximately 50 ms later, the normal force vanishes completely, and the bar enters free flight. Leveraging the impulse-momentum principle, the bar's rotation is accelerated by the pivoting velocity.
• Figure 4: Schematic of the two forward models to predict the landing pose of a new command $\mathbf{\hat{u}}$.
• Figure 5: Initial population of landing poses generated from impulse-momentum-based control using a mesh of $3\times3\times3=27$ commands. Each command is executed 5 times, yielding 135 throws.