deep-rl-grasping
deep-rl-grasping copied to clipboard
BarisYazici
Train on custom objects
Your work is very interesting. How can I train for new custom objects or test your trained model on new objects?
Yes, you can do both training and testing on custom objects. I provided one different wooden block object set under model folder. You need to give the config file the name of this dataset. In this case, wooden blocks and those models will be spawned in the scene. For your custom dataset, you also need to change the base scene object sampler and add the function to sample from your custom object dataset. If it is similar to the pybullet random-urdfs object dataset you can use a similar function to this to partition the dataset to train-val-test.
Thank you for your fast answering. I have changed in "deep-rl-grasping/config/gripper_grasp.yaml" but nothing is changed. Then I have changed the "deep-rl-grasping/trained_models/SAC_depth_1mbuffer/best_model/config.yaml" and then I saw the wooden blocks for testing. I will test for custom objects as well and let you know about the results. Thanks again.
Thank you for your interesting work. I am writing description for others to apply other datasets.
For custom objects I have used This repository to make SDF file from YCB dataset. Then I have added This code to base_scene.py
def _sample_ycb(self, n_objects): self._model_path = "models/" object_names = [ '011_banana'] selection = self._rng.choice(object_names, size=n_objects) paths = [os.path.join(self._model_path, 'ycb'+'/'+str(name), name + '.sdf') for name in selection] return paths, 1.
and added
object_samplers = {'ycb': self._sample_ycb, 'wooden_blocks': self._sample_wooden_blocks, 'random_urdfs': self._sample_random_objects}
So if you change the data_set: ycb in any using config.yaml then you have YCB objects in the scene.
I want to mention that for scaling objects you can change the number in return paths, scale in def_sample_ycb function
Note: The urdf file name should be the same with object_names
Perfect! I am glad that you work it out. Let me know about the test results :)
@BarisYazici Could you help me to run the code with Robot Kuka? As I run in KukaExt branch, there is also only gripper without robot. I want to know exact changes that I should do to bring whole robot.
@BarisYazici Second question about command ' pytest tests_gripper' when I try to run it, I get this errors:
~/RL/deep-rl-grasping-kukaExt$ pytest tests_gripper ================================================ test session starts ================================================= platform linux -- Python 3.6.13, pytest-6.2.2, py-1.10.0, pluggy-0.13.1 rootdir: /home/lab/RL/deep-rl-grasping-kukaExt collected 0 items / 1 error
======================================================= ERRORS =======================================================
_____________________________________ ERROR collecting tests_gripper/test_sim.py _____________________________________
tests_gripper/test_sim.py:17: in
@BarisYazici Could you help me to run the code with Robot Kuka? As I run in KukaExt branch, there is also only gripper without robot. I want to know exact changes that I should do to bring whole robot.
You need to change the robot model path under the config file
@BarisYazici Second question about command ' pytest tests_gripper' when I try to run it, I get this errors:
~/RL/deep-rl-grasping-kukaExt$ pytest tests_gripper ================================================ test session starts ================================================= platform linux -- Python 3.6.13, pytest-6.2.2, py-1.10.0, pluggy-0.13.1 rootdir: /home/lab/RL/deep-rl-grasping-kukaExt collected 0 items / 1 error
======================================================= ERRORS ======================================================= _____________________________________ ERROR collecting tests_gripper/test_sim.py _____________________________________ tests_gripper/test_sim.py:17: in env_encoder_simp = gym.make('gripper-env-v0', config=CONFIG_ENCODER_SIMP) ../../miniconda3/envs/grasp_env/lib/python3.6/site-packages/gym/envs/registration.py:145: in make return registry.make(id, **kwargs) ../../miniconda3/envs/grasp_env/lib/python3.6/site-packages/gym/envs/registration.py:90: in make env = spec.make(kwargs) ../../miniconda3/envs/grasp_env/lib/python3.6/site-packages/gym/envs/registration.py:60: in make env = cls(_kwargs) manipulation_main/gripperEnv/robot.py:110: in init config, self._camera, self) manipulation_main/gripperEnv/sensor.py:192: in init self._encoder.load_weights(model_dir) manipulation_main/gripperEnv/encoders.py:30: in load_weights self._model.load_weights(weights_path) ../../miniconda3/envs/grasp_env/lib/python3.6/site-packages/keras/engine/network.py:1166: in load_weights f, self.layers, reshape=reshape) ../../miniconda3/envs/grasp_env/lib/python3.6/site-packages/keras/engine/saving.py:1004: in load_weights_from_hdf5_group original_keras_version = f.attrs['keras_version'].decode('utf8') E AttributeError: 'str' object has no attribute 'decode'
Sorry the tests only cover the master branch. I didn't write tests for Kuka branch. But if you are interested, go ahead and do it. You can submit a pull request
@BarisYazici Could you help me to run the code with Robot Kuka? As I run in KukaExt branch, there is also only gripper without robot. I want to know exact changes that I should do to bring whole robot.
You need to change the robot model path under the config file
It doesn't work.
@BarisYazici Second question about command ' pytest tests_gripper' when I try to run it, I get this errors: ~/RL/deep-rl-grasping-kukaExt$ pytest tests_gripper ================================================ test session starts ================================================= platform linux -- Python 3.6.13, pytest-6.2.2, py-1.10.0, pluggy-0.13.1 rootdir: /home/lab/RL/deep-rl-grasping-kukaExt collected 0 items / 1 error ======================================================= ERRORS ======================================================= _____________________________________ ERROR collecting tests_gripper/test_sim.py _____________________________________ tests_gripper/test_sim.py:17: in env_encoder_simp = gym.make('gripper-env-v0', config=CONFIG_ENCODER_SIMP) ../../miniconda3/envs/grasp_env/lib/python3.6/site-packages/gym/envs/registration.py:145: in make return registry.make(id, **kwargs) ../../miniconda3/envs/grasp_env/lib/python3.6/site-packages/gym/envs/registration.py:90: in make env = spec.make(kwargs) ../../miniconda3/envs/grasp_env/lib/python3.6/site-packages/gym/envs/registration.py:60: in make env = cls(_kwargs) manipulation_main/gripperEnv/robot.py:110: in init config, self._camera, self) manipulation_main/gripperEnv/sensor.py:192: in init self._encoder.load_weights(model_dir) manipulation_main/gripperEnv/encoders.py:30: in load_weights self._model.load_weights(weights_path) ../../miniconda3/envs/grasp_env/lib/python3.6/site-packages/keras/engine/network.py:1166: in load_weights f, self.layers, reshape=reshape) ../../miniconda3/envs/grasp_env/lib/python3.6/site-packages/keras/engine/saving.py:1004: in load_weights_from_hdf5_group original_keras_version = f.attrs['keras_version'].decode('utf8') E AttributeError: 'str' object has no attribute 'decode'
Sorry the tests only cover the master branch. I didn't write tests for Kuka branch. But if you are interested, go ahead and do it. You can submit a pull request
I got the same error in master branch.
@BarisYazici Could you help me to run the code with Robot Kuka? As I run in KukaExt branch, there is also only gripper without robot. I want to know exact changes that I should do to bring whole robot.
You need to change the robot model path under the config file
It doesn't work.
Do you have the same problem? Can you post the error messages when the simulation dies?
@BarisYazici Could you help me to run the code with Robot Kuka? As I run in KukaExt branch, there is also only gripper without robot. I want to know exact changes that I should do to bring whole robot.
You need to change the robot model path under the config file
It doesn't work.
Do you have the same problem? Can you post the error messages when the simulation dies?
It doesn't die, just the robot doesn't loaded when I change the config file which you mentioned file while in Kuka branch the config file is written for kuka but by the way it is not loaded and just the gripper only is loaded.
@BarisYazici Could you help me to run the code with Robot Kuka? As I run in KukaExt branch, there is also only gripper without robot. I want to know exact changes that I should do to bring whole robot.
You need to change the robot model path under the config file
It doesn't work.
Do you have the same problem? Can you post the error messages when the simulation dies?
It doesn't die, just the robot doesn't loaded when I change the config file which you mentioned file while in Kuka branch the config file is written for kuka but by the way it is not loaded and just the gripper only is loaded.
If you are trying to run a trained model then you need to change the specific config file under that model file. For example here at this file. For every model in the trained folder directory, best_model folder if you want to run that one, you need to change the respective config file under the best_model folder.
@BarisYazici Could you help me to run the code with Robot Kuka? As I run in KukaExt branch, there is also only gripper without robot. I want to know exact changes that I should do to bring whole robot.
You need to change the robot model path under the config file
It doesn't work.
Do you have the same problem? Can you post the error messages when the simulation dies?
It doesn't die, just the robot doesn't loaded when I change the config file which you mentioned file while in Kuka branch the config file is written for kuka but by the way it is not loaded and just the gripper only is loaded.
If you are trying to run a trained model then you need to change the specific config file under that model file. For example here at this file. For every model in the trained folder directory, best_model folder if you want to run that one, you need to change the respective config file under the best_model folder.
The result is as follows:
That seems like the wrong model of the file. The kuka model I use has another gripper. It should be this sdf model. Can you share with me your config.yaml file settings?
It was kuka_with_gripper2.sdf but for kuka_with_gripper.sdf also the same. The config file is like this:
`BDQ: batch_size: 256 buffer_size: 10000 epsilon_greedy: true exploration_final_eps: 0.02 exploration_fraction: 0.1 layers:
- [64, 64]
- [32]
- [32] learning_rate: 0.00169477235301313 learning_starts: 1000 num_actions_pad: 16 prioritized_replay: true save_dir: BDQ1mSimplified target_network_update_freq: 1000 tensorboard_logs: null total_timesteps: 5000000 train_freq: 300 DDPG: {save_dir: DDPG, tensorboard_logs: DDPG, total_timesteps: 2000000} DQN: {batch_size: 32, learning_rate: 0.001, prioritized_replay: true, save_dir: DQN4mFull, tensorboard_logs: null, total_timesteps: 4000000} PPO: {learning_rate: 0.001, n_steps: 2000, save_dir: ppo_5m, tensorboard_logs: tensorboard_logs/ppo_5m, total_timesteps: 1000000} SAC: batch_size: 64 buffer_size: 1000000 layers: [64, 64] max_iters: 400 save_dir: SAC_full_rgbd step_size: 0.0003 tensorboard_logs: null total_timesteps: 5000000 TRPO: {batch_size: 64, max_iters: 400, save_dir: trpo_1m, step_size: 0.001, tensorboard_logs: tensorboard_logs/trpo_1m, total_timesteps: 5000000} algorithm: sac curriculum: extent: [0.01, 0.1] init_lambda: 0.0 lift_dist: [0.01, 0.1] max_objects: [3, 5] min_objects: [1, 1] n_steps: 8 robot_height: [0.15, 0.25] success_threshold: 0.7 window_size: 1000 depth_observation: false discount_factor: 0.99 full_observation: true normalize: true reward: {custom: true, delta_z_scale: 1000.0, grasp_reward: 100.0, shaped: true, terminal_reward: 10000.0, time_penalty: 200.0} #models/gripper/wsg50_one_motor_gripper_new.sdf, robot: {discrete: false, max_force: 100, max_translation: 0.03, max_yaw_rotation: 0.15, model_path: models/kuka_iiwa/kuka_with_gripper.sdf , num_actions_pad: 2, step_size: 0.01, yaw_step: 0.1} scene: {data_set: random_urdfs, scene_type: OnFloor} sensor: camera_info: config/camera_info.yaml encoder_dir: encoder_files/new_gripper_encoder randomize: {focal_length: 4, optical_center: 2, rotation: 0.0349, translation: 0.002} transform: config/camera_transform.yaml visualize: false simplified: false simulation: {real_time: false, visualize: false} skip_empty_initial_state: true time_horizon: 150`