evelyd/paper_delia_2025_iros_physics-informed_trajectory_generation_dataset

Stabilizing Humanoid Robot Trajectory Generation via Physics-Informed Learning and Control-Informed Steering

Evelyn D'Elia, Paolo Maria Viceconte, Lorenzo Rapetti, Diego Ferigo, Giulio Romualdi, Giuseppe L'Erario, Raffaello Camoriano, and Daniele Pucci

Submitted to the 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).

📂 Dataset

The dataset is organized in folders. The layout is explained below.

• The mocap/D2 folder contains the raw motion capture data, acquired via a human using the Xsens motion tracking suit. Within D2 there are 5 subfolders for each data subset: forward, backward, side, diagonal, and mixed walking. each subfolder contains a log file which can be retargeted onto the robot.
• The retargeted_mocap folder contains the text files of the retargeted mocap data, in both original (D2) and mirrored (D2_mirrored) form.
• The IO_features folder contains the text files of the extracted features (data and labels) to be used in training. There are 4 subfolders which correspond to the original and mirrored inputs and the original and mirrored outputs.
• The onnx folder contains the onnx files corresponding to trained models from each PI loss value as well as from the original ADHERENT framework.
• The reproduce_plots folder contains .mat files with all the data needed to reproduce the selected plots.