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

iros_2025_submission_video_github.mp4

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

Installation | Paper | Dataset

Installation

We recommend using conda/mamba, available here, for this installation.

First, create and activate a conda environment with the necessary dependencies.

mamba create -n pi_trajectory_gen -c conda-forge bipedal-locomotion-framework jaxsim pytorch tensorboard adam-robotics jax2torch urdf-parser-py h5py ergocub-models meshcat-python
mamba activate pi_trajectory_gen

Next, within the conda environment, clone and install this repo:

git clone /~https://github.com/ami-iit/paper_delia_2025_iros_physics-informed_trajectory_generation.git
cd paper_delia_2025_iros_physics-informed_trajectory_generation
pip install .

You will also need to download the datasets necessary for running the code into this repo. This may take some time depending on your internet connection.

git clone git@hf.co:datasets/evelyd/paper_delia_2025_iros_physics-informed_trajectory_generation_dataset datasets/
cd datasets/
unzip D2.zip

Now you're ready to run the code.

Running the code

To replicate our results, several scripts are available in the scripts folder.

cd ../scripts/

Retargeting the data

The raw dataset contains 5 subsets: forward walking, backward walking, side walking, diagonal walking, and mixed walking. The raw data is recorded on a human and therefore needs to be retargeted onto the robot model. For example, the mirrored version of the forward walking subset is retargeted using:

python retargeting.py --KFWBGR --filename ../datasets/D2/1_forward_normal_step/data.log

Extracting features from the data

With the retargeted data we can extract the features (data and labels) for training, for example:

python features_extraction.py --dataset D2 --portion 1

Training the model

With the extracted features we can train the model. The weight given to the PI loss component can be specified as an argument.

python training.py --pi_weight 1.0

Running the trajectory generator with a trained model

With a trained model, we can give dummy inputs and generate a trajectory based on the model's predictions.

python trajectory_generation.py

The various parameters used for the trajectory generation can be tuned in the files present in the config folder. For example, in config_mann.toml, the trained model can be changed by editing the onnx_model_path parameter, and the correction block gains can be updated by changing the linear_pid_gain and rotational_pid_gain parameters.

Replicating plots in the paper

We also include a script to reproduce some of the plots from our paper.

python show_selected_plots.py

Maintainer

@evelyd