GINGA: Humanoid Sports Dataset

José G. R. Teles^1*†, Juan R. B. de F. Silva^2*†, Lavínia R. Lima^3*†, Lucas Ólives^3*†, Marcos P. C. M. Queiroz ^4*†, Telma W. de L. Soares ^4*†

¹Institute of Physics, Federal University of Goiás, Goiânia, Brazil

²Division of Electrical and Computer Engineering, Technological Institute of Aeronautics, São José dos Campos, Brazil

³School of Electrical, Mechanical and Computer Engineering, Federal University of Goiás, Goiânia, Brazil

⁴Institute of Informatics, Federal University of Goiás, Goiânia, Brazil

^*Alphabetical order, equal contribution

^†AKCIT: Advanced Knowledge Center for Immersive Technologies

Motivation

The RoboCup initiative posits a grand challenge: by the mid-21st century, a team of fully autonomous humanoid robot soccer players shall play a soccer game against the winner of the most recent World Cup, complying with official FIFA rules. To bridge the gap between current robotic capabilities and human-level performance, robots must transcend basic walking gaits. They require "GINGA", a term synonymous with the Brazilian spirit of movement, representing agility, rhythm, and creativity. Currently, the development of athletic behaviors in humanoids faces a significant bottleneck regarding data. While large-scale datasets exist, they often lack specific utility for competitive sports. A robot goalkeeper needs to perform a defense instantly, while a striker needs to execute different types of shots to score goals and celebrate to engage with the crowd. Existing datasets often lack these specific, culturally rich, and physically explosive motions.

Evaluation

Data

We utilized an Xsens inertial motion capture system to capture high-fidelity joint rotations without the occlusion issues common in optical systems. The dataset is organized into two distinct motion categories:

• Game Actions: Functional movements required for competitive play, including single-leg balances, sprinting, step-overs, and the Trivela kick;
• Human-Robot Interaction (HRI) movements: Motions crucial for communication and social connection, featuring routines like the CR7 Siu celebration, the Macarena, and Neymar's Paradinha Penalti.

During the capture process, several challenges inherent to inertial technology were addressed to ensure high fidelity:

• We performed physical checks and initial calibrations for each sensor, as incorrect positioning generated significant distortions in the kinematic model;
• Because the software's biomechanical solver is highly sensitive to the actor's dimensions, we created unique anthropometric profiles for each participant;
• To combat the accumulation of error (drift) caused by prolonged, explosive movements, we periodically interrupted sessions to recalibrate the entire suit;
• A careful post-processing step was implemented to rotate and align the initial pose of all clips to the same global reference frame.

Method

Our methodology provides a robust workflow to transfer human motion to a physical humanoid. A critical challenge in this domain is the kinematic mismatch between the human actor and the robot. To resolve this, we employed General Motion Retargeting (GMR) to map the Xsens skeletal data to the URDF/MJCF description of the Unitree G1 robot, ensuring physical constraints were respected prior to training.

To validate the utility of the Ginga dataset, we trained a control policy using Reinforcement Learning via the Beyond Mimic framework in IsaacLab. Our approach incorporated specific training mechanisms:

• An adaptive sampling mechanism identified frames where the robot frequently lost balance and increased their sampling frequency, accelerating convergence;
• A compact tracking reward combined local joint imitation with world-frame global constraints to reduce long-term drift in locomotion tasks;
• We decoupled the control frequency from the native capture rate (240 FPS) and extended episode truncation limits to safely process longer gestures and respect the robot's physical constraints.

Results

The training metrics indicate that the Unitree G1 robot agent was able to successfully track the kinematic references of all categories in the dataset. Average reward curves for movements, such as the featured celebrations, showed a rapid and consistent rise in the first 5,000 steps, reaching a plateau of stability between 140 and 190 reward points. The variation in the total asymptotic performance is a direct consequence of the different durations of the original reference clips, rather than a reflection of tracking failures or policy instability.

The robustness of these policies was verified through cross-validation between simulators, where the training performed in IsaacLab was successfully tested on MuJoCo. The movements occurred fluidly, replicating the human athlete's intention despite the challenges of inertial technology, such as the absence of ground reaction force data.

Related Work

High-fidelity motion data is foundational for data-driven animation, though its quality is linked to acquisition hardware. Optical systems have been the industry standard for repositories like AMASS and CMU, but they are often constrained to studios and suffer from occlusions. Inertial systems provide an occlusion-free alternative crucial for capturing explosive sports dynamics, yet datasets utilizing them for competitive soccer remain scarce.

Physics-based character control has evolved from trajectory tracking, as seen in DeepMimic, to utilizing deep reinforcement learning and adversarial priors (AMP) to encourage natural motion styles. To apply these learned policies to diverse robot morphologies, retargeting techniques like GMR are essential to handle structural discrepancies. Our methodology builds upon these foundations by utilizing the Beyond Mimic framework, emphasizing the robust recovery required for transferring dynamic motions to the Unitree G1.

Furthermore, the domain of humanoid sports has shifted toward learning-based agility, with significant advancements in bipedal fall recovery, agile ball manipulation, and extreme balance tasks. While these works achieve impressive functional competence, they lack the variety of movements found in actual gameplay and human interaction. Our work complements this existing research by introducing GINGA, integrating culturally rich celebrations alongside competitive sports movements.