Learning Symmetric and Low-energy Locomotion

This is code for our paper: https://arxiv.org/abs/1801.08093

Install from script

The script below has been tested on clean Ubuntu 16.04/18.04 computers.

sudo -s
source prereq_install.sh
exit

Manual Install

The code consists of two parts: dart-env, which is an extension of OpenAI Gym that uses Dart for rigid-body simulation, and baselines, which is adapted from OpenAI Baselines. Both libraries are included in the repository so you will not need to install the original library.

To install dart-env, you need to install Dart and Pydart, follow the instructions at: https://github.com/DartEnv/dart-env/wiki for installing these two packages. After installing Dart and Pydart, install dart-env by:

cd dart-env
pip install -e .

To install baselines, execute the following:

cd baselines
pip install -e .

How to use

Example scripts for training and testing the policies are provided in siggraph_script. The specific python code corresponding to each script can be found here.

The training results will be saved to data/. The final policy is saved as policy_params.pkl. You can also find the intermediate policies in the folders organized by the corresponding curriculums. To test a policy, run:

python baselines/test_policy.py ENV_NAME PATH_TO_POLICY

To visualize the learning curve, run:

python baselines/plot_benchmark.py PATH_TO_FOLDER

Setup environment

4 example environments are included: DartWalker3d-v1, DartHumanWalker-v1, DartDogRobot-v1 and DartHexapod-v1.

The desired velocity is controlled by three variables in the initialization of each environment: init_tv sets the target velocity at the beginning of the rollout, final_tv sets the target velocity we want the character to reach eventually, and tv_endtime sets the amount of time (in seconds) it takes to accelerate from init_tv to final_tv.

Setup training script

Refer to run_walker3d_staged_learning.py for an example on how to setup the training script for the biped walking robot.

Mirror-symmetry

The mirror-symmetry loss for a new environment is configured with the argument observation_permutation and action_permutation when initializing MlpMirrorPolicy in the training script.

For observation_permutation and action_permutation, they are two vectors used for mirror symmetric loss. Each entry in these two denotes the index of the corresponding entry in observation/action AFTER it is mirrored w.r.t the sagittal plane of the character, and the sign of the element means whether the entry should multiply -1 after mirroring. For example, if a character has its left and right elbow joint angle at index 4 and 7 of the obsevation vector, then observation_permutation[4] should be 7 and observation_permutation[7] should be 4. Further, if the behavior of the two dofs are opposite, e.g. larger value of left elbow angle means flexion while larger value of right elbow angle means extension, then a -1 should be multiplied to both entries in observation_permutation. Note that for dofs at the center of the character (like pelvis rotation), their corresponding mirrored entry are simply themselves, with -1 multiplied to some of them. Also, if the entry at index 0 need to be negated, you need to use a small negative value like -0.0001, as multiplying -1 wouldn't change 0.

Additional notes

For a newly created dart-env environment, you can use examine_skel.py to test the model configurations, which I found to be helpful in debugging joint limits.

Refer to run_walker3d_staged_learning.py for an example on how to setup the training script for the biped walking robot.

ODE Internal Error

If you see errors like: ODE INTERNAL ERROR 1: assertion "d[i] != dReal(0.0)" failed in _dLDLTRemove(), try downloading lcp.cpp and replace the one in dart/external/odelcpsolver/ with it. Recompile Dart and Pydart2 afterward and the issue should be gone.