概要

若想训练自己的场景，数据集的重要性不做过多赘述，下面就基于 robomimic 和 robosuite 构建自己的数据集进行讲解，同时，也会附上 train 和 run 的流程，这样，就形成了闭环。

自建数据集

采集数据

采集数据可使用脚本 collect_human_demonstrations.py 完成,在采集过程中,需要自己定义 env 的相关信息,在实际使用时,存在以下几个问题：

• 无法控制机器人精准的完成抓取工作
• 机器人在某些姿态下,运动会出现漂移的情况

格式转换

该功能脚本为 convert_robosuite.py,在进行执行是:

$ python conversion/convert_robosuite.py --dataset /path/to/demo.hdf5

生成OBS用于训练

python dataset_states_to_obs.py --dataset demo_myliftdraw.hdf5 --output AAA.hdf5

BUG记录

Bug1

在进行 conver_to_robosuite时，也有个小bug，当数据集中 demo数量较少时，比如为1，那么 convert 就会失败，这是因为在进行 split 时数据太少导致的，

"""
Helper script to convert a dataset collected using robosuite into an hdf5 compatible with
this repository. Takes a dataset path corresponding to the demo.hdf5 file containing the
demonstrations. It modifies the dataset in-place. By default, the script also creates a
90-10 train-validation split.For more information on collecting datasets with robosuite, see the code link and documentation
link below.Code: https://github.com/ARISE-Initiative/robosuite/blob/offline_study/robosuite/scripts/collect_human_demonstrations.pyDocumentation: https://robosuite.ai/docs/algorithms/demonstrations.htmlExample usage:python convert_robosuite.py --dataset /path/to/your/demo.hdf5
"""import h5py
import json
import argparseimport robomimic.envs.env_base as EB
from robomimic.scripts.split_train_val import split_train_val_from_hdf5if __name__ == "__main__":parser = argparse.ArgumentParser()parser.add_argument("--dataset",type=str,help="path to input hdf5 dataset",)args = parser.parse_args()f = h5py.File(args.dataset, "a") # edit mode# import ipdb; ipdb.set_trace()# store env metaenv_name = f["data"].attrs["env"]env_info = json.loads(f["data"].attrs["env_info"])env_meta = dict(type=EB.EnvType.ROBOSUITE_TYPE,env_name=env_name,env_version=f["data"].attrs["repository_version"],env_kwargs=env_info,)if "env_args" in f["data"].attrs:del f["data"].attrs["env_args"]f["data"].attrs["env_args"] = json.dumps(env_meta, indent=4)print("====== Stored env meta ======")print(f["data"].attrs["env_args"])# store metadata about number of samplestotal_samples = 0for ep in f["data"]:# ensure model-xml is in per-episode metadataassert "model_file" in f["data/{}".format(ep)].attrs# add "num_samples" into per-episode metadataif "num_samples" in f["data/{}".format(ep)].attrs:del f["data/{}".format(ep)].attrs["num_samples"]n_sample = f["data/{}/actions".format(ep)].shape[0]f["data/{}".format(ep)].attrs["num_samples"] = n_sampletotal_samples += n_sample# add total samples to global metadataif "total" in f["data"].attrs:del f["data"].attrs["total"]f["data"].attrs["total"] = total_samplesf.close()# create 90-10 train-validation split in the datasetsplit_train_val_from_hdf5(hdf5_path=args.dataset, val_ratio=0.1)

最后一行，只需要将 val_ration = 1 就可以了。

Bug2

描述:同样的操作过程,用自己新建的 env环境类生成数据,在进行 obs 生成时报错,但是用官方的 env 是没有问题的

解决方法：不要试图随性构建自己的模型库，因为这玩意是按照固定路径搜索的，坑坑坑，忙了半天，发现是路径的问题，里面有个 robosuite 的固定关键词，这个会影响 to obs 的。

Train && Test

Train 需要调用的脚本是：

$ python robomimic/robomimic/scripts/train.py --config bc_rnn.json

Test调用的脚本为：

$ python robomimic/robomimic/scripts/un_trained_agent.py --agent /path/to/model.pth --n_rollouts 50 --horizon 400 --seed 0 --video_path /path/to/output.mp4 --camera_names agentview robot0_eye_in_hand 

在进行训练时，需要对算法进行配置，即导入的 json 文件，以下是一个配置文件的案例，该文件来源于 mimicgen 生成的，其中，里面 有对于算法的部分参数， 比如 "algo_name": "bc",如果功底不够深厚的话，建议大家只改下数据集的 path 就可以了:

{"algo_name": "bc","experiment": {"name": "source_draw_low_dim","validate": false,"logging": {"terminal_output_to_txt": true,"log_tb": true,"log_wandb": false,"wandb_proj_name": "debug"},"save": {"enabled": true,"every_n_seconds": null,"every_n_epochs": 50,"epochs": [],"on_best_validation": false,"on_best_rollout_return": false,"on_best_rollout_success_rate": true},"epoch_every_n_steps": 100,"validation_epoch_every_n_steps": 10,"env": null,"additional_envs": null,"render": false,"render_video": true,"keep_all_videos": false,"video_skip": 5,"rollout": {"enabled": true,"n": 50,"horizon": 400,"rate": 50,"warmstart": 0,"terminate_on_success": true}},"train": {"data": "/home/idm/Documents/mimicgen/robosuite/robosuite/scripts/generate_dataset/tmp/draw_data/demo_obs.hdf5","output_dir": "/home/idm/Documents/mimicgen/robosuite/robosuite/scripts/generate_dataset/tmp/draw_data/trained_models","num_data_workers": 0,"hdf5_cache_mode": "all","hdf5_use_swmr": true,"hdf5_load_next_obs": false,"hdf5_normalize_obs": false,"hdf5_filter_key": null,"hdf5_validation_filter_key": null,"seq_length": 10,"pad_seq_length": true,"frame_stack": 1,"pad_frame_stack": true,"dataset_keys": ["actions","rewards","dones"],"goal_mode": null,"cuda": true,"batch_size": 100,"num_epochs": 2000,"seed": 1},"algo": {"optim_params": {"policy": {"optimizer_type": "adam","learning_rate": {"initial": 0.001,"decay_factor": 0.1,"epoch_schedule": [],"scheduler_type": "multistep"},"regularization": {"L2": 0.0}}},"loss": {"l2_weight": 1.0,"l1_weight": 0.0,"cos_weight": 0.0},"actor_layer_dims": [],"gaussian": {"enabled": false,"fixed_std": false,"init_std": 0.1,"min_std": 0.01,"std_activation": "softplus","low_noise_eval": true},"gmm": {"enabled": true,"num_modes": 5,"min_std": 0.0001,"std_activation": "softplus","low_noise_eval": true},"vae": {"enabled": false,"latent_dim": 14,"latent_clip": null,"kl_weight": 1.0,"decoder": {"is_conditioned": true,"reconstruction_sum_across_elements": false},"prior": {"learn": false,"is_conditioned": false,"use_gmm": false,"gmm_num_modes": 10,"gmm_learn_weights": false,"use_categorical": false,"categorical_dim": 10,"categorical_gumbel_softmax_hard": false,"categorical_init_temp": 1.0,"categorical_temp_anneal_step": 0.001,"categorical_min_temp": 0.3},"encoder_layer_dims": [300,400],"decoder_layer_dims": [300,400],"prior_layer_dims": [300,400]},"rnn": {"enabled": true,"horizon": 10,"hidden_dim": 400,"rnn_type": "LSTM","num_layers": 2,"open_loop": false,"kwargs": {"bidirectional": false}},"transformer": {"enabled": false,"context_length": 10,"embed_dim": 512,"num_layers": 6,"num_heads": 8,"emb_dropout": 0.1,"attn_dropout": 0.1,"block_output_dropout": 0.1,"sinusoidal_embedding": false,"activation": "gelu","supervise_all_steps": false,"nn_parameter_for_timesteps": true}},"observation": {"modalities": {"obs": {"low_dim": ["robot0_eef_pos","robot0_eef_quat","robot0_gripper_qpos","object"],"rgb": [],"depth": [],"scan": []},"goal": {"low_dim": [],"rgb": [],"depth": [],"scan": []}},"encoder": {"low_dim": {"core_class": null,"core_kwargs": {},"obs_randomizer_class": null,"obs_randomizer_kwargs": {}},"rgb": {"core_class": "VisualCore","core_kwargs": {},"obs_randomizer_class": null,"obs_randomizer_kwargs": {}},"depth": {"core_class": "VisualCore","core_kwargs": {},"obs_randomizer_class": null,"obs_randomizer_kwargs": {}},"scan": {"core_class": "ScanCore","core_kwargs": {},"obs_randomizer_class": null,"obs_randomizer_kwargs": {}}}},"meta": {"hp_base_config_file": null,"hp_keys": [],"hp_values": []}
}

部分训练截图如下，可以看出，因数据太少，导致模型不收敛：

参考资料

RoboMimic