This repository contains the code and the neural networks used for our paper "Learning Time-optimized Path Tracking with or without Sensory Feedback".
The code is written in python and does not need to be compiled. Simply clone the repository with
git clone https://github.com/translearn/pathTracking.git
The required dependencies can be installed by running:
pip install -r requirements.txt
We provide pretrained networks for the robot systems shown in the figure above. To track paths from a random dataset with an industrial robot run
python tracking/evaluate.py --use_gui --checkpoint=industrial/no_balancing/random
Other networks can be executed by adjusting the checkpoint argument. All available networks are listed below:
| Robot system | Configuration | Dataset | Checkpoint |
|---|---|---|---|
| Kuka iiwa | no additional objectives | random | --checkpoint=industrial/no_balancing/random |
| target point | --checkpoint=industrial/no_balancing/target_point | ||
| ball balancing | --checkpoint=industrial/no_balancing/ball_balancing | ||
| Kuka with balance board | no balancing reward | ball balancing | --checkpoint=industrial/balancing/no_balancing_reward |
| balancing reward | --checkpoint=industrial/balancing/balancing_reward | ||
| ARMAR-6 | no additional objectives | random | --checkpoint=humanoid/armar6/random |
| ARMAR-4 | no additional objectives, fixed base and legs | random | --checkpoint=humanoid/armar4/no_balancing/random |
| target point | --checkpoint=humanoid/armar4/no_balancing/target_point | ||
| no balancing reward, fixed legs | target point | --checkpoint=humanoid/armar4/balancing/no_balancing_reward | |
| balancing reward, fixed legs | --checkpoint=humanoid/armar4/balancing/balancing_reward_fixed_legs | ||
| balancing reward, controlled legs | --checkpoint=humanoid/armar4/balancing/balancing_reward_controlled_legs |
Networks can also be trained from scratch. For instance, path tracking with an industrial robot can be learned by running
python tracking/train.py --logdir=tracking_training --name=industrial_no_balancing_random --robot_scene=0 --online_trajectory_time_step=0.1 --hidden_layer_activation=swish --online_trajectory_time_step=0.1 --online_trajectory_duration=16.0 --obstacle_scene=0 --target_link_offset="[0, 0, 0.126]" --last_layer_activation=tanh --no_log_std_activation --use_controller_target_velocities --spline_dir=industrial/random/train --spline_u_arc_start_range="[0.0, 0.8]" --spline_u_arc_diff_min=0.2 --spline_normalize_duration --spline_termination_max_deviation=0.25 --obs_spline_n_next=7 --obs_spline_add_length --obs_spline_add_distance_per_knot --spline_distance_max_reward=2.0 --spline_deviation_max_threshold=0.25 --punish_spline_max_deviation --spline_max_deviation_max_punishment=0.9 --punish_spline_mean_deviation --spline_mean_deviation_max_punishment=0.9 --spline_deviation_weighting_factors="[1.0, 1.0, 1.0, 0.9, 0.8, 0.7, 0.6]" --batch_size_factor=6.0 --spline_braking_extra_time_steps=0 --terminate_on_robot_stop --solver_iterations=50 --iterations_per_checkpoint=50 --time=500The corresponding publication is available at https://arxiv.org/abs/2203.01968.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.