Transferable Dynamics Learning

Benchmarking of different dynamics learning methods under changes in the distribution over controls. In this repository we provide the evaluation code and links to the corresponding evaluation data sets.

Robotic setup

The data was recorded on a 3-DOF torque-controlled real-robotic system. Please check the Open Dynamic Robot Initiative for the details about the robotic platform.

Robotic dataset

All the dataset files can be downloaded from here or alternatively from here

We use the aforementioned robotic platform to record different datasets under substantially different conditions in order to assess the predictive performance of dynamics learning algorithms beyond the iid. setting. In this study such conditions are in turn governed by the particular controllers used to generate the control inputs feed into the system. We consider two types of controllers depending on whether there is a feedback control loop or not.

Closed-loop dataset - Sine waves PD control

This dataset was generated using a superposition of sine waves to generate trajectories that were subsequently tracked by the robot using PD position control (Please refer to the paper for details). Check out some sampled robot movements under the 4 different families of controllers considered in the closed-loop dataset:

The shown movements are arranged according to the following diagram where each of the datasets D account for a particular configuration of sine angular frenquencies (low, high) and reachable task space (left, full). The arrows denote the different transfer settings that are discussed in our paper.

Open-loop dataset - GP torque controls

We used sampled trajectories from a Gaussian process (GP) directly as torque inputs (controls) to record this dataset in an open loop fashion. Please refer to the paper for details.

Dataset structure and naming convention.

The released files of the closed-loop dataset are named according to the following convention:

• Sines_full.npz
• Sines_training.npz
• Sines_validation.npz
• Sines_test_iid.npz
• Sines_test_transfer_%d.npz

with the different test transfer files being indexed starting from 1 (e.g., Sines_test_transfer_1.npz). Note that each _full.npz files includes all recorded rollouts of its corresponding category. The rest of the files can be obtained from it as follows:

python -m DL.utils.data_extractor --raw_data Sines_full.npz --training Sines_training.npz --testiid Sines_test_iid.npz --validation Sines_validation.npz --testtransfer Sines_test_transfer_{}.npz

The open-loop dataset files are equally named, except that the prefix GPs is used instead of Sines.

Each of the released npz files can be indexed by the following keywords which account for different recorded variables:

• measured_angles
• measured_velocities
• measured_torques
• constrained_torques
• desired_torques

Each of these keys is associated with a numpy array of shape (S, T, D), where S denotes the number of sequences/rollouts, and T, D are the sequence length and number of degrees of freedom, respectively. We recorded at a frequency of 1 Hz, meaning that the time elapsed between consecutive observations is 0.001 seconds. We recorded using a 3-DOF finger robot (D=3), and all rollouts have a total duration of 14s (T=14000).

Simulated closed-loop dataset

We also provide a simulated version of the closed-loop dataset. We keep the naming convention consistent with the real datasets but use the prefix Sim_Sines instead.

Evaluation.

Our evaluation protocol consists of two stages. First, we compute the error vectors of a particular method over different datasets and save them in a single .npz file. After we compute the error files corresponding to all benchmarked methods we aggregate the results in different plots as shown in the paper, which we refer to for further details.

Computing error vectors

As an example the following command computes the aforementioned error file for a linear model trained with SGD:

python -m DL.evaluation.evaluation \
--method linear_model_sgd \
--training_data Sines_training.npz \
--validation_data Sines_validation.npz \
--iid_test_data Sines_test_iid.npz \
--transfer_test_data Sines_test_transfer_1.npz Sines_test_transfer_2.npz Sines_test_transfer_3.npz \
--prediction_horizon 1 \
--history_length 1 \
--no-averaging \
--streaming \
--verbose \
--output_errors errors.npz

Among the implemented methods we have Gaussian Processes (SVGPR), System Identification methods (system_id), Neural networks NN, etc.

Dependencies

All dependencies are tracked using Pipenv. In order to reproduce our Python environment with all dependencies type the following command in this project's directory (transferable_dynamics_dataset):

pipenv install

After this the virtual environment can be activated by typing:

pipenv shell

Paper & Reference

Preprint.

In order to cite us please do so according to:

@conference{AgudeloEspanaetal20,
  title = {A Real-Robot Dataset for Assessing Transferability of Learned Dynamics Models },
  author = {Agudelo-España, D. and Zadaianchuk, A. and Wenk, P. and Garg, A. and Akpo, J. and Grimminger, F. and Viereck, J. and Naveau, M. and Righetti, L. and Martius, G. and Krause, A. and Sch{\"o}lkopf, B. and Bauer, S. and W{\"u}thrich, M.},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year = {2020}
}