Mastering Agile Jumping Skills from Simple Practices with Iterative Learning Control

TL;DR

The paper addresses the challenge of achieving precise target jumping for legged robots amid long aerial phases and contact-uncertainty, which can risk hardware damage. It introduces an Iterative Learning Control framework that starts from simple jumps and progressively handles harder targets, guided by a motor-dynamic constraint model and a three-stage optimization to improve safety and accuracy. The approach is validated both in simulation and on hardware (Unitree A1), demonstrating cm-level position and near 0–1 degree orientation accuracy within a few trials, even with unknown payloads or variable ground contacts. By leveraging knowledge from simple tasks, the method enables data-efficient learning for agile jumping and shows promise for generalization to other legged platforms and 3D maneuvers.

Abstract

Achieving precise target jumping with legged robots poses a significant challenge due to the long flight phase and the uncertainties inherent in contact dynamics and hardware. Forcefully attempting these agile motions on hardware could result in severe failures and potential damage. Motivated by these challenging problems, we propose an Iterative Learning Control (ILC) approach that aims to learn and refine jumping skills from easy to difficult, instead of directly learning these challenging tasks. We verify that learning from simplicity can enhance safety and target jumping accuracy over trials. Compared to other ILC approaches for legged locomotion, our method can tackle the problem of a long flight phase where control input is not available. In addition, our approach allows the robot to apply what it learns from a simple jumping task to accomplish more challenging tasks within a few trials directly in hardware, instead of learning from scratch. We validate the method via extensive experiments in the A1 model and hardware for various jumping tasks. Starting from a small jump (e.g., a forward leap of 40cm), our learning approach empowers the robot to accomplish a variety of challenging targets, including jumping onto a 20cm high box, jumping to a greater distance of up to 60cm, as well as performing jumps while carrying an unknown payload of 2kg. Our framework can allow the robot to reach the desired position and orientation targets with approximate errors of 1cm and 1 degree within a few trials.

Figures (12)

• Figure 1: Learn and practice from easy to more challenging tasks. Our approach enables the A1 robot to apply what it learned from a simple jump of $40~cm$ in order to accomplish more challenging tasks: (a) Jump farther to $60~cm$ within 9 trials, (b) jump on box at $(x,z)=(60, 10)~cm$ within 8 trials, (c) Jump on box at $(x,z)=(50, 20)~cm$ within 7 trials. Supplemental video: https://youtu.be/zbEB5bMBgY0
• Figure 2: Practicing a simple task. This framework describes the learning process of a simple jumping maneuver
• Figure 3: Learning to complete further challenging tasks. Our proposed framework enables the learning from a simple task to more challenging goals within several trials, instead of learning from scratch. Joint reference profile of simple task $\bm{q}_d, \dot{\bm{q}}_d$ can be utilized for challenging tasks instead of re-running the trajectory optimization to get a new jumping reference
• Figure 4: Jump from soft ground. The figures show the comparison between model-free PD-type ILC, baseline ILC-MPC, and our proposed ILC in terms of body trajectories and body orientation for the jumping forward to a target at $0.6~m$ from soft ground with $(K_p^G, K_d^G)=(2.10^3, 5.10^2)$.
• Figure 5: Jump from hard ground. The plots show the learning progress over trials with model-free PD-type ILC, baseline ILC-MPC, and our proposed ILC when the robot jumping from the hard ground with $(K_p^G, K_d^G)=(2.10^4, 3.10^3)$ .