Learning Collision-free and Torque-limited Robot Trajectories based on Alternative Safe Behaviors

This repository provides the code to learn torque-limited and collision-free robot trajectories without exceeding limits on the position, velocity, acceleration and jerk of each robot joint.

Installation

To use the code, clone the repository with:

git clone https://github.com/translearn/safeMotions.git

The required dependencies can be installed by running:

pip install -r requirements.txt

Trajectory generation  

To generate a random trajectory with a single robot run

python safemotions/random_agent.py --use_gui --check_braking_trajectory_collisions --check_braking_trajectory_torque_limits --torque_limit_factor=0.6 --plot_trajectory

For a demonstration scenario with two robots run

python safemotions/random_agent.py --use_gui --check_braking_trajectory_collisions --robot_scene=1

Collision-free trajectories for three robots can be generated by running

python safemotions/random_agent.py --use_gui --check_braking_trajectory_collisions --robot_scene=2

Pretrained networks  

Various pretrained networks for reaching randomly sampled target points are provided.
Make sure you use ray==1.4.1 to open the pretrained networks.

Industrial robots

To generate and plot trajectories for a reaching task with a single industrial robot run

python safemotions/evaluate.py --use_gui --plot_trajectory --plot_actual_torques --checkpoint=industrial/one_robot/collision 

Trajectories for two and three industrial robots with alternating target points can be generated by running

python safemotions/evaluate.py --use_gui --checkpoint=industrial/two_robots/collision/alternating  

and

python safemotions/evaluate.py --use_gui --checkpoint=industrial/three_robots/collision/alternating  

Humanoid robots

	ARMAR 6	ARMAR 6x4
Alternating target points	--checkpoint=humanoid/armar6/collision/alternating	--checkpoint=humanoid/armar6_x4/collision/alternating
Simultaneous target points	--checkpoint=humanoid/armar6/collision/simultaneous	--checkpoint=humanoid/armar6_x4/collision/simultaneous
Single target point	--checkpoint=humanoid/armar6/collision/single	--checkpoint=humanoid/armar6_x4/collision/single

Training

Networks can also be trained from scratch. For instance, a reaching task with a single robot can be learned by running

python safemotions/train.py --logdir=safemotions_training --name=industrial_one_robot_collision --robot_scene=0 --online_trajectory_time_step=0.1 --hidden_layer_activation=swish --online_trajectory_duration=8.0 --obstacle_scene=3 --use_target_points --target_point_sequence=0 --target_point_cartesian_range_scene=0 --target_link_offset="[0, 0, 0.126]" --target_point_radius=0.065 --obs_add_target_point_pos --obs_add_target_point_relative_pos --check_braking_trajectory_collisions --closest_point_safety_distance=0.01 --acc_limit_factor_braking=1.0 --jerk_limit_factor_braking=1.0 --punish_action --action_punishment_min_threshold=0.95 --action_max_punishment=0.4  --target_point_reached_reward_bonus=5  --pos_limit_factor=1.0 --vel_limit_factor=1.0 --acc_limit_factor=1.0 --jerk_limit_factor=1.0 --torque_limit_factor=1.0 --punish_braking_trajectory_min_distance --braking_trajectory_min_distance_max_threshold=0.05 --braking_trajectory_max_punishment=0.5 --last_layer_activation=tanh --solver_iterations=50 --normalize_reward_to_initial_target_point_distance --collision_check_time=0.033 --iterations_per_checkpoint=50 --time=200

Publication

The corresponding publication is available at https://arxiv.org/abs/2103.03793.

Disclaimer

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.