add realman shadow src

realman_src/realman_aloha/shadow_rm_act/config/config.yaml

@@ -0,0 +1,74 @@
+robot_env: {
+  # TODO change the path to the correct one
+  rm_left_arm: '/home/rm/aloha/shadow_rm_aloha/config/rm_left_arm.yaml',
+  rm_right_arm: '/home/rm/aloha/shadow_rm_aloha/config/rm_right_arm.yaml',
+  arm_axis: 6,
+  head_camera: '215222076892',
+  bottom_camera: '215222076981',
+  left_camera: '152122078151',
+  right_camera: '152122073489',
+  # init_left_arm_angle: [0.226, 21.180, 91.304, -0.515, 67.486, 2.374, 0.9],
+  # init_right_arm_angle: [-1.056, 33.057, 84.376, -0.204, 66.357, -3.236, 0.9]
+  init_left_arm_angle: [6.45, 66.093, 2.9, 20.919, -1.491, 100.756, 18.808, 0.617],
+  init_right_arm_angle: [166.953, -33.575, -163.917, 73.3, -9.581, 69.51, 0.876]
+}
+dataset_dir: '/home/rm/aloha/shadow_rm_aloha/data/dataset/20250103'
+checkpoint_dir: '/home/rm/aloha/shadow_rm_act/data'
+# checkpoint_name: 'policy_best.ckpt'
+checkpoint_name: 'policy_9500.ckpt'
+state_dim: 14
+save_episode: True
+num_rollouts: 50                #训练期间要收集的 rollout（轨迹）数量
+real_robot: True
+policy_class: 'ACT'
+onscreen_render: False
+camera_names: ['cam_high', 'cam_low', 'cam_left', 'cam_right']
+episode_len: 300        #episode 的最大长度（时间步数）。
+task_name: 'aloha_01_11.28'
+temporal_agg: False     #是否使用时间聚合
+batch_size: 8         #训练期间每批的样本数。
+seed: 1000            #随机种子。
+chunk_size: 30      #用于处理序列的块大小
+eval_every: 1           #每隔 eval_every 步评估一次模型。
+num_steps: 10000        #训练的总步数。
+validate_every: 1       #每隔 validate_every 步验证一次模型。
+save_every: 500         #每隔 save_every 步保存一次检查点。
+load_pretrain: False     #是否加载预训练模型。
+resume_ckpt_path: 
+name_filter:  # TODO
+skip_mirrored_data: False    #是否跳过镜像数据（例如用于基于对称性的数据增强）。
+stats_dir:  
+sample_weights:  
+train_ratio: 0.8       #用于训练的数据比例（其余数据用于验证）
+
+policy_config: {
+  hidden_dim: 512, # Size of the embeddings (dimension of the transformer)
+  state_dim: 14, # Dimension of the state
+  position_embedding: 'sine', # ('sine', 'learned').Type of positional embedding to use on top of the image features
+  lr_backbone: 1.0e-5,
+  masks: False, # If true, the model masks the non-visible pixels
+  backbone: 'resnet18',
+  dilation: False, # If true, we replace stride with dilation in the last convolutional block (DC5)
+  dropout: 0.1, # Dropout applied in the transformer
+  nheads: 8,
+  dim_feedforward: 3200, # Intermediate size of the feedforward layers in the transformer blocks
+  enc_layers: 4, # Number of encoding layers in the transformer
+  dec_layers: 7, # Number of decoding layers in the transformer
+  pre_norm: False, # If true, apply LayerNorm to the input instead of the output of the MultiheadAttention and FeedForward
+  num_queries: 30,
+  camera_names: ['cam_high', 'cam_low', 'cam_left', 'cam_right'],
+  vq: False,
+  vq_class: none,
+  vq_dim: 64,
+  action_dim: 14,
+  no_encoder: False,
+  lr: 1.0e-5,
+  weight_decay: 1.0e-4,
+  kl_weight: 10,
+
+  # lr_drop: 200,
+  # clip_max_norm: 0.1,
+}
+
+
+