A Kung Fu Athlete Bot That Can Do It All Day: Highly Dynamic, Balance-Challenging Motion Dataset and Autonomous Fall-Resilient Tracking

TL;DR

The paper tackles the lack of high-dynamic motion data and robust autonomous fall recovery for humanoid robots by introducing KungFuAthlete, a dataset of ground and jump martial-arts motions derived from professional training videos. It couples a data-correction pipeline (root-height correction and Savitzky–Golay smoothing) with an end-to-end policy trained via FastSAC that unifies high-dynamic motion tracking and fall recovery, aided by gravity-based state initializations and low-kinetic-energy sampling. Key contributions include the KungFuAthlete dataset, a comprehensive root-height correction framework, a unified end-to-end learning objective for tracking and recovery, and extensive simulation-to-real validation on a Unitree G1 with domain randomization, supported by ablation studies on reward design. The approach advances robust, autonomous humanoid performance in high-dynamic environments, enabling safer operation beyond controlled labs and cages.

Abstract

Current humanoid motion tracking systems can execute routine and moderately dynamic behaviors, yet significant gaps remain near hardware performance limits and algorithmic robustness boundaries. Martial arts represent an extreme case of highly dynamic human motion, characterized by rapid center-of-mass shifts, complex coordination, and abrupt posture transitions. However, datasets tailored to such high-intensity scenarios remain scarce. To address this gap, we construct KungFuAthlete, a high-dynamic martial arts motion dataset derived from professional athletes' daily training videos. The dataset includes ground and jump subsets covering representative complex motion patterns. The jump subset exhibits substantially higher joint, linear, and angular velocities compared to commonly used datasets such as LAFAN1, PHUMA, and AMASS, indicating significantly increased motion intensity and complexity. Importantly, even professional athletes may fail during highly dynamic movements. Similarly, humanoid robots are prone to instability and falls under external disturbances or execution errors. Most prior work assumes motion execution remains within safe states and lacks a unified strategy for modeling unsafe states and enabling reliable autonomous recovery. We propose a novel training paradigm that enables a single policy to jointly learn high-dynamic motion tracking and fall recovery, unifying agile execution and stabilization within one framework. This framework expands robotic capability from pure motion tracking to recovery-enabled execution, promoting more robust and autonomous humanoid performance in real-world high-dynamic scenarios.

Figures (8)

• Figure 1: Teaser of the KungFuAthlete dataset. Visualization of reconstructed full-body motions from daily training videos of professional martial artists. The dataset spans a wide spectrum of motion dynamics, from non-jumping ground actions to highly dynamic acrobatic jumps, resulting in significantly larger joint velocities and body-level dynamics than existing human motion datasets. This diversity makes KungFuAthlete particularly suitable for pushing the boundaries of humanoid motion and exploring the current limits of robotic capabilities.
• Figure 2: Overall experimental demonstration.The red arrows indicate the direction and magnitude of the pulling force on the torso, with longer lines representing greater force.
• Figure 3: Illustration of the height data correction process.
• Figure 4: Corrected Root Node Curve Diagram.
• Figure 5: Flowchart of training a policy using FastSAC for high-dynamic tracking and autonomous fall recovery without reference frames under low-kinetic-energy sampling.