Skip to content

Latest commit

 

History

History
78 lines (65 loc) · 3.2 KB

README.md

File metadata and controls

78 lines (65 loc) · 3.2 KB
Raw

Lin_graduation

Content

The repository structure is adapted from gym-pybullet-drones . The main training, evaluation and visualization code files are under folder /experiments/learning.

  • 1000_data contains the 1000 success/failure scenario data for visualization.
  • csv contains the tensorboard log.
  • model_logs contains the final trained policies.

Folder name 08.30_18.08_t_ppo_opt90_25_1_adam_cpu___FIN is constructed with date(08.30_18.08)_environment(t)_algorithm(ppo)_note(opt90_25_1_adam_cpu).

The addition note opt90_25_1_adam_cpu stands for 90% optimization rate; 25 for PPO batch size; 1 for PPO epoch number; adam optimizer and trained on CPU.

Example

The training, evaluation and visualization code files are under folder /experiments/learning.

Data generation

The training and testing dataset can be generated with

python3 a0_dataset_gen.py

Training

The libraries used in this project are: stable-baselines3 and stable-baselines3-contrib.

  • a0_training_w_dataset.py is for policy training in the default two-obstacle-environment.
$ python3 a0_training_w_dataset.py --it 500 --note trail_training

with arguments

  • --exp: define the direction of a policy for continue training.
  • --it: define the training iteration (1 iteration = n_steps).
  • --note: add addition note to the policy folder.
  • --vel: set the observation for history obstacle position to obstacle velocity.
After calling the training file, the algorithm selection is presented:
Please select the algorithm
 0) PPO; 1) A2C; 2) TRPO; 3) RPPO (seperate indexes with comma)

Typing in 0,0,0,1,1,1 represents 3 trainings for PPO and 3 training for A2C.

Evaluation

  • a0_model_eva.py is for the success rate evaluation and trajectory visualization.
$ python3 a0_model_eva.py --exp ./simple_model_logs/08.30_18.08_t_ppo_opt90_25_1_adam_cpu___FIN --model 13200 --subs

with arguments

  • --exp: define the direction of the policy.
  • --model: define which policy to evaluate. Default setting is success; best or iteration number (like 13200) can also be used.
  • --env: define which environment to load. Default environment is two-obstacle-environment; orbit or act is also available.
  • --subs: define whether to plot the subplot figure for multiple scenarios.
  • --it: define the amount of test scenario for success rate computation.

Choose the policy to evaluate via:

Please type the algorithm index to continue training
(0, '0_PPO')
(1, '0_PPO1')
(2, '0_PPO2')
(3, '1_A2C')
(4, '1_A2C1')
(5, '1_A2C2')
(seperate indexes with comma or type 'all')

Animation

The animation code is adapted from gym-pybullet-drones . The urdf models are under folder /gym_pybullet_drones/assets

  • a0_animation.py is for generating trajectory animation.
$ python3 a0_animation.py --exp ./simple_model_logs/08.30_18.08_t_ppo_opt90_25_1_adam_cpu___FIN/0_PPO1/subs_log

a0_animation.py reads the eva_result_{}.npy file under argument --exp