3D-Diffusion-Policy
2025-04-24
25julyfunnotespolicy符号T: 预测轨迹长度eval_policy.py 调用 deploy_policy.py:eval()，其调用 RobotRunner.get_action()RobotRunner.get_action() (robot_runner.py)obs = self.get_n_steps_obs()obs <- update_obs() 就是 append <- Base_task.get_obs()[ai]observation - 包含来自各种相机的观察数据相机数据包括 head_camera, left_camera, right_camera, front_camera 等每个相机可以包含以下数据（取决于 data_type 设置）：rgb - RGB图像数据mesh_segmentation - 网格分割数据actor_segmentation - 实体分割数据depth - 深度图像数据相机内参和外参矩阵pointcloud - 点云数据如果 data_type['pointcloud'] 为 True，则包含点云数据可以选择是否合并多个相机的点云数据joint_action - 机器人关节状态如果 data_type['qpos'] 为 True，包含机器人关节角度双臂模式下，包含左臂和右臂的关节角度单臂模式下，仅包含右臂的关节角度endpose - 机器人末端执行器姿态如果 data_type['endpose'] 为 True，包含末端执行器的位置和姿态双臂模式下，包含左右两个末端执行器的信息（位置x,y,z，欧拉角roll,pitch,yaw，以及夹爪状态）单臂模式下，仅包含右臂末端执行器信息vision_tactile - 视觉触觉传感器数据（当 TACTILE_ON 为 True 时）如果 data_type['vision_tactile'] 为 True，包含触觉传感器的RGB图像数据随后拿出两个数据并重命名: pointcloud -> point_cloud, joint_action -> agent_pos得到 obs: Dicteach_key => 将最近 n 个观测的 key 在第 0 维度拼接. 形状为 (n_steps, ) + shape_of_the_valuen_steps 在参数 yaml 里为 n_obs_steps = 3在前面 unsqueeze 一个长度为 1 的维度后送进 DP3.predict_action() （应该是因为推理的时候 batch 必是 1）class DP3:
predict_action() and (dp3.py)
[arg] (obs_dict):
'point_cloud': (1, 3, 1024, 6)
'agent_pos': (3, 14) 就是关节角度
normalize()
if @?global_cond:
point_cloud & agent_pos 都送入 DP3Encoder，得到 (3, 192)，压扁成 (1, 576)
mask 就是全部 mask 掉, 所有动作都需要通过扩散模型生成
else:
mask 观察特征保持可见
送入 self.condition_sapmle()
return. 实测表明一次预测 6 步且会把这 6 步执行完，再预测下 6 步.
...
action_pred: (B, T, action_dim) e.g. (1, 8, 14)
start = To - 1
end = start + self.n_action_steps
action = action_pred[:, start:end] e.g. (1, 6, 14)
condition_sample():
生成的 traj shape 是 (B, T, action_dim) = (1, 8, 14)
每个去噪步 model(sample=trajectory, timestep=t, local_cond=local_cond(必为 None), global_cond=global_cond)
model is @ConditionalUnet1D.forward():[ConditionalUnet1D.forward()]:
...
timestep: (形状 (B, ) or int)
encoding: (SinusoidalPosEmb, Linear, Mish, Linear)
if global_cond: global_feature = cat([timestep_embed, global_cond], axis=-1) }
[Downsample]
x = sample (生成 trajactory)
for idx, (@resnet, @resnet2, @downsample) in enumerate(self.down_modules):
if self.use_down_condition:
x = resnet(x, global_feature)
if idx == 0 and len(h_local) > 0:
x = x + h_local[0]
x = resnet2(x, global_feature)
else:
x = resnet(x)
if idx == 0 and len(h_local) > 0:
x = x + h_local[0]
x = resnet2(x)
h.append(x)
x = downsample(x)
[mid_module (@ConditionalResidualBlock1D)]
...
[Upsample]
对称的
[return]
x = self.final_conv(x)
return x
[resnet] = @ConditionalResidualBlock1D(
dim_in, dim_out, cond_dim=cond_dim,
kernel_size=kernel_size, n_groups=n_groups,
condition_type=condition_type
),
[resnet2] = ConditionalResidualBlock1D(
dim_out, dim_out, cond_dim=cond_dim,
kernel_size=kernel_size, n_groups=n_groups,
condition_type=condition_type
),
[downsample] = Downsample1d(dim_out) if not is_last else nn.Identity()
[ConditionalResidualBlock1D].forward():
out = Conv1dBlock() (x)
if `cross_attention_add`
embed = CrossAttention() (x)
out = out + embed (是 tensor 值加)
out = another Conv1dBlock() (x)
out = out + self.residual_conv(x)
return outDP3Encoder:
'point_cloud' => B x N x 3的 点云 (B: batch) = (3, 1024, 3) 送入 PointNetEncoderXYZ:
class PointNetEncoderXYZ
MLP: [Linear + LayerNorm + ReLU] x 3
channels: 3 => 64 => 128 => 256 => max => Linear + LayerNorm (128)
forward():
(B, N, 3) = (3, 1024, 3)
mlp => (3, 1024, 256)
max => (3, 256)
Linear + LayerNorm => (3, 128)
self.state_mlp: 简单的 MLP (Linear + ReLU). state_mlp_size = (64, 64).
最后 cat 成 (3, 192