Save the dirty bit to the chunk directory.

Saves an MP4 replay of an episode.

Load or create a task configuration and set up robot embodiments. Args: task_name (str): Task name. Returns: tuple: (Ta

Converts all numpy arrays in nested dictionary or list or tuple to torch tensors (and leaves existing torch Tensors as-is), and returns a

Build multimodal embeddings and attention mask

Get the mtime of a given date at midnight UTC.

Used to convert between aloha and pi joint angles.

Used to convert between aloha and pi joint angles.

Helper method to prepare 4D attention masks for transformer.

Helper to get action masks from labels

Set folder permission to be read, write and searchable.

Set folder permission to be read, write and searchable.

chmod requires executable permission to be set, so we skip if the permission is already match with the target.

chmod requires executable permission to be set, so we skip if the permission is already match with the target.

Add a marker for the centroid. Returns: Updated scene with centroid marker

Add arrows pointing from each face center to the centroid. Returns: Updated scene with face arrows

Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors (https://qwenlm.github.io/blog/qwen2-vl/). Explanation:

Apply `functional` variant of TIMM's Transform = Compose([Resize -> CenterCrop -> ToTensor -> Normalize])

Converts gripper actions from continuous to binary values (0 and 1). We exploit that fact that most of the time, the gripper is fully open (

Capitalize the first letter of a string and add a period to the end if it's not there.

将单张图片压缩为 JPEG 格式的二进制数据。 参数: img (numpy.ndarray): 输入的图像数组，形状为 (height, width, channels)。 返回: compressed_data (bytes): 压缩

This long function is unfortunately doing something very simple and is being very defensive: We are separating out all parameters of

1x1 convolution

3x3 convolution with padding

Counts and prints the number of trainable parameters in a module. Args: module (nn.Module): PyTorch module. module_name (str

Create a table with specified dimensions.

Enrich the actor list by adding 'functional_points' and 'contact_points' from the corresponding model metadata file, and remove the 'modelnam

TASK_ENV: Task Environment Class, you can use this class to interact with the environment model: The model from 'get_model()' function ob

Do a full training forward pass and compute the loss (batch_size x num_steps x num_motors)

Generate descriptions for episodes by replacing placeholders in instructions with parameter values. For each episode, filter instructions tha

Generate the json dict and state for an episode.

For each dt window, choose the first timestamp in the window. Assumes timestamps sorted. One timestamp might be chosen multiple times due to

Get action head for continuous value prediction. Args: cfg: Configuration object with model parameters llm_dim: Dimension of

获取从 v1 到 v2 的旋转矩阵

Get a conversation template.

Get image resize dimensions for a specific model. If returned value is an int, the resized image will be a square. If returned value is

Generate position embeddings for multimodal conditions. mm_cond_lens: an OrderedDict containing (modality name, modality token

获取物体应当被放置到的位置 考虑因素： 1. 三维坐标与给定坐标一致 2. 物体的朝向合理 - 物体 z 轴与给定坐标 z 轴一致 - 满足在 xy 平面上的一定约束 -

Get the VLA model's Hugging Face processor. Args: cfg: Configuration object with model parameters Returns: AutoProcesso

Get proprioception projector for the VLA model. Args: cfg: Configuration object with model parameters llm_dim: Dimension of