Whole-Body Dynamic Throwing with Legged Manipulators

TL;DR

The paper addresses the challenge of throwing with legged manipulators by enabling full-body, 3D throws that leverage body momentum while maintaining stability. It introduces a DPPO-based RL framework trained in IsaacLab with an adaptive curriculum to balance accuracy and stability, formalized by a compact reward $r = \lambda_1 r_{\text{throwing}} + \lambda_2 r_{\text{stability}} + \lambda_3 r_{\text{roll}}$ and a 3D target error $E = \min_{t} \| x_t y_t z_t - x_A y_A z_A \|_2$. The method is evaluated on a humanoid and an armed quadruped, showing superior full-body performance over arm-only approaches, generalization to arbitrary 3D targets, and sim-to-real transfer with real-world throws achieving around $0.73$–$0.80$ m error. Key contributions include a novel full-body throwing formulation, an adaptive curriculum design to mitigate local minima, and comprehensive cross-morphology evaluation with hardware validation.

Abstract

Throwing with a legged robot involves precise coordination of object manipulation and locomotion - crucial for advanced real-world interactions. Most research focuses on either manipulation or locomotion, with minimal exploration of tasks requiring both. This work investigates leveraging all available motors (full-body) over arm-only throwing in legged manipulators. We frame the task as a deep reinforcement learning (RL) objective, optimising throwing accuracy towards any user-commanded target destination and the robot's stability. Evaluations on a humanoid and an armed quadruped in simulation show that full-body throwing improves range, accuracy, and stability by exploiting body momentum, counter-balancing, and full-body dynamics. We introduce an optimised adaptive curriculum to balance throwing accuracy and stability, along with a tailored RL environment setup for efficient learning in sparse-reward conditions. Unlike prior work, our approach generalises to targets in 3D space. We transfer our learned controllers from simulation to a real humanoid platform.

Figures (4)

• Figure 1: Our robot throwing policies demonstrated on real hardware (top) and in simulation (bottom) showing complex full-body movements that improve throwing accuracy and distance, while maintaining a stable configuration. The top figure shows our humanoid setup and a long-range throw deployed sim2real and the real ball trajectory overlayed and highlighted in orange measured with motion capture.
• Figure 2: Throwing error (metres) between robot throw and target for the humanoid (left) and the quadruped (right) in simulation experiments. Shaded region represents $\pm$ 1 standard deviation over 5 runs. PyTorch random seed was randomised only for humanoid experiments.
• Figure 3: Policy Evaluation and Real-world deployment architecture.
• Figure 4: Throwing error difference (full-body policy$-$arm-only policy) from the target (metres) for the humanoid (left) and quadruped (right), at a distance of 5 and 8 metres respectively. Error difference displayed over all discretised throwing angles ($\phi$,$\theta$). Averaged over 3 trials per angle. Red indicates higher performance from full-body.