Open X-Embodiment（最大规模之一）
- 来自 21+ 机构、22 种不同机器人（单臂、双臂、四足等）的 1M+ 真实机器人轨迹，统一格式（RLDS）。包含图像、动作、语言指令等，多样性极高，被称为"机器人版的 ImageNet"。
- 用途 ：训练跨 embodiment（不同机器人身体）的通用策略（X-robot learning）、行为克隆、世界模型等。促进零样本/少样本迁移。
Open X-Embodiment 中的数据组织方式是：
- Dataset → 由很多 Episodes（一条完整的演示轨迹）组成
- Episode → 由很多 Steps（时间步，每一步对应一帧）组成
- 每个 Step 包含下面这些核心字段
单个step展示：
python('action', {'open_gripper': <tf.Tensor: shape=(), dtype=bool, numpy=True>, 'rotation_delta': <tf.Tensor: shape=(3,), dtype=float32, numpy=array([ 6.077167e-07, -1.193009e-07, 1.308389e-07], dtype=float32)>, 'terminate_episode': <tf.Tensor: shape=(), dtype=float32, numpy=0.0>, 'world_vector': <tf.Tensor: shape=(3,), dtype=float32, numpy=array([1.9514002e-10, 8.0674190e-11, 2.9859176e-10], dtype=float32)>}) ('is_first', <tf.Tensor: shape=(), dtype=bool, numpy=True>) ('is_last', <tf.Tensor: shape=(), dtype=bool, numpy=False>) ('is_terminal', <tf.Tensor: shape=(), dtype=bool, numpy=False>) ('observation', {'image': <tf.Tensor: shape=(480, 640, 3), dtype=uint8, numpy= array([[[ 83, 88, 91], [ 92, 97, 100], [ 95, 100, 103], ..., [ 12, 16, 19], [ 11, 15, 18], [ 11, 15, 18]], [[ 99, 104, 107], [112, 117, 120], [ 77, 82, 85], ..., [ 12, 16, 19], [ 12, 16, 19], [ 11, 15, 18]], [[109, 114, 117], [102, 107, 110], [ 73, 78, 81], ..., [ 13, 17, 20], [ 12, 15, 20], [ 12, 15, 20]], ..., [[209, 199, 189], [212, 202, 192], [210, 200, 190], ..., [159, 120, 89], [159, 120, 89], [160, 121, 88]], [[214, 197, 187], [211, 197, 186], [212, 199, 190], ..., [156, 120, 88], [156, 120, 88], [157, 121, 89]], [[214, 196, 186], [217, 200, 190], [215, 201, 192], ..., [155, 118, 89], [155, 119, 87], [156, 120, 88]]], dtype=uint8)>, 'natural_language_embedding': <tf.Tensor: shape=(512,), dtype=float32, numpy= array([ 2.30294447e-02, 6.05173744e-02, -5.20481169e-02, -2.47768629e-02, 1.08427750e-02, 9.11148265e-03, -5.82255470e-03, -5.61384659e-04, -6.19610846e-02, 2.86639649e-02, -5.38054272e-04, 3.64868939e-02, 5.83116375e-02, -6.58905273e-03, -8.61448571e-02, -4.80751926e-03, -1.01660937e-02, 1.18107673e-04, -2.88839266e-02, 3.15110236e-02, 2.88712289e-02, -3.36049236e-02, -3.32037471e-02, 8.23714286e-02, -4.78581823e-02, 1.89173147e-02, -1.17043108e-02, 2.34112926e-02, 7.34530063e-03, -1.72176119e-02, 2.04275567e-02, 3.81773487e-02, 4.12922464e-02, 4.79167514e-02, -3.39735448e-02, 3.30767594e-02, -4.11522724e-02, -1.53829195e-02, -5.84172420e-02, 3.31984945e-02, 5.99250644e-02, 6.65826425e-02, 8.60797986e-03, -5.16836308e-02, -1.56743526e-02, 1.52221220e-02, -1.06724147e-02, 4.62612659e-02, -6.90690577e-02, -1.53337680e-02, -4.61382791e-03, -5.53948805e-02, -5.04228892e-03, -8.98943655e-03, -6.67139981e-03, 3.24231014e-02, 2.31685992e-02, 6.53641894e-02, -8.42219740e-02, -1.27285672e-03, -8.77480879e-02, 1.38156675e-02, -1.46904064e-03, 1.15498211e-02, 5.21797240e-02, -5.19705354e-04, -2.24918984e-02, 6.42215684e-02, -4.64617684e-02, 2.62698326e-02, 3.75501104e-02, -3.29499803e-02, -7.24997371e-02, -8.21422189e-02, -1.41956005e-02, 3.98824103e-02, 1.40000479e-02, 8.12648833e-02, -2.12381762e-02, -2.07397826e-02, -5.43150529e-02, 3.39643732e-02, -1.70414969e-02, -2.75766440e-02, 2.39402782e-02, 5.05681075e-02, 2.82037798e-02, 4.60241176e-02, 2.80120317e-03, -5.19922487e-02, 6.25411570e-02, 4.05185223e-02, -1.27093755e-02, -1.11003146e-02, 6.20534718e-02, 1.30008266e-03, 5.72689734e-02, 5.93838990e-02, 1.60258804e-02, -5.33626489e-02, -2.78736539e-02, 5.59482542e-05, 6.45361319e-02, 3.67150083e-02, 4.29617465e-02, 8.05545747e-02, 1.70597266e-02, 1.83145087e-02, -2.10689139e-02, 1.55236656e-02, 6.38596108e-03, 9.61762015e-03, -3.68559025e-02, -5.65365329e-02, -8.20576102e-02, 4.36546106e-04, 1.85061377e-02, 1.79954898e-02, -4.18148264e-02, -1.34398844e-02, -5.50495237e-02, -5.81721216e-03, -3.58555908e-03, 8.61355010e-03, -1.84907299e-02, 4.18909490e-02, -7.99810067e-02, 4.61478308e-02, -1.94716249e-02, 1.08241506e-01, 4.07698080e-02, 8.40406120e-02, 1.12400670e-02, -1.70313604e-02, 6.71370700e-02, 3.07028089e-03, -7.81752318e-02, -3.61328013e-02, 9.47262254e-03, -9.52577963e-03, 5.29839359e-02, 1.38539104e-02, -2.02670484e-03, -4.78461385e-02, -2.62118988e-02, -4.16636690e-02, 5.06601445e-02, -9.39371157e-03, -4.18891832e-02, 1.33636931e-03, -3.27013247e-02, -4.38975617e-02, 3.90706435e-02, -3.39448899e-02, -1.37101451e-03, 1.01938769e-01, 5.05746491e-02, 2.22387109e-02, -9.20815207e-03, -5.59622720e-02, 2.36338340e-02, 3.57812233e-02, 3.67419459e-02, 4.05951738e-02, -3.91846336e-02, -3.26995626e-02, 6.80791363e-02, 7.62440078e-03, 3.71609740e-02, 1.11309215e-02, 2.43490692e-02, -2.58231871e-02, -4.75365296e-03, -1.32513493e-02, 4.95327450e-02, -2.94215400e-02, 1.48535492e-02, 4.26678807e-02, 5.19396178e-02, 2.48162784e-02, -3.35754268e-02, 1.53736938e-02, -5.17860018e-02, -1.55840889e-02, -1.37480170e-01, 1.45085938e-02, 4.23489325e-02, -3.09551191e-02, -6.17873259e-02, 1.18616760e-01, 1.93447457e-03, -3.51321213e-02, 3.47951464e-02, 1.00223430e-01, -5.11648208e-02, -3.18400338e-02, -3.42703797e-02, 1.03139855e-01, 3.10028344e-02, 6.91802651e-02, 6.74038455e-02, -2.69975960e-02, 9.42844898e-03, 2.41422597e-02, -1.10973073e-02, -8.41486901e-02, 3.67963091e-02, 3.02007329e-02, 5.21205217e-02, -1.60790365e-02, 4.82646599e-02, 5.85531294e-02, 5.83571829e-02, 2.33832877e-02, 1.15526440e-02, 3.20059396e-02, 1.97310932e-02, 2.08420889e-03, -3.63293551e-02, -3.13506350e-02, 8.75868872e-02, 2.91291904e-02, 9.95325111e-03, 8.91545340e-02, 4.88411672e-02, 2.94628982e-02, 5.47678173e-02, -2.45558042e-02, -4.52916101e-02, 7.55258650e-02, 1.61027964e-02, -1.17763421e-02, -4.52294350e-02, 4.45150882e-02, 3.60127762e-02, -8.18681996e-03, 7.46363848e-02, -1.79726221e-02, 4.47655469e-02, 8.11863169e-02, 3.53159979e-02, -1.39076794e-02, 3.07287946e-02, -2.10502259e-02, -8.53858422e-03, 2.70146951e-02, -4.04026592e-03, 6.43077912e-03, 1.19745508e-02, 2.06797030e-02, -8.06324333e-02, -8.93529970e-04, 3.63447075e-03, 2.07282994e-02, -5.89092001e-02, 6.57288581e-02, -6.81186020e-02, 9.78687033e-03, 2.74124667e-02, 7.90238082e-02, -1.72475800e-02, -7.87314959e-03, -7.10541308e-02, -1.38236294e-02, 7.51328692e-02, 4.54298928e-02, 1.40015250e-02, 2.35893158e-03, -5.71069727e-03, -7.58356750e-02, 1.15318755e-02, -9.97513458e-02, -7.44394362e-02, 1.94973976e-03, -5.39304875e-02, -3.12894508e-02, 5.76846413e-02, -1.61580462e-02, 1.20309796e-02, 2.00632177e-02, -4.42018919e-02, -9.56523046e-03, -1.34884613e-02, 4.89868447e-02, -5.09263668e-03, -4.25534621e-02, 2.55978890e-02, -5.81441559e-02, 9.62353572e-02, 2.02117823e-02, -1.71392057e-02, -8.66092928e-03, -7.12014129e-03, 2.09330823e-02, 5.37754819e-02, -2.28657983e-02, -3.00242249e-02, -2.37361230e-02, 5.92173003e-02, 9.49474890e-03, -4.23376933e-02, 8.67886376e-03, -2.29124147e-02, 4.76709800e-03, -6.53485348e-03, 3.07827536e-02, -1.12981182e-02, -6.45409599e-02, -4.22426425e-02, -4.35791314e-02, -3.22807357e-02, 6.40643388e-02, 2.14385260e-02, 6.49540052e-02, -2.70295776e-02, 7.27179274e-02, 7.42838776e-04, -1.51439067e-02, -7.70827755e-02, 5.93844475e-03, -5.38883060e-02, -8.49346817e-02, 6.47232234e-02, 2.66256160e-03, -6.04295097e-02, 3.50991786e-02, 6.68003969e-03, 5.09417169e-02, 5.67321070e-02, -6.97394609e-02, -2.94399057e-02, -2.46931631e-02, -1.35134598e-02, -1.35464333e-02, -8.00555572e-02, -4.59090434e-02, 2.26265122e-03, -2.62019224e-02, -5.06492890e-02, 1.75190102e-02, -6.88601807e-02, -1.22997751e-02, 7.56917149e-02, 6.47063181e-02, -5.43286018e-02, 2.50871237e-02, -3.73206660e-02, -2.14841217e-02, 6.80395355e-03, 2.31319037e-03, -5.60809672e-02, 1.24006355e-02, -5.31012006e-02, -3.00859232e-02, -4.90449194e-04, 7.92119950e-02, -5.58438292e-03, -9.47208703e-02, 3.93651873e-02, -9.36861802e-03, -2.10053977e-02, 4.35312726e-02, -4.43299953e-03, 8.10990185e-02, -2.13196576e-02, 2.98558734e-02, -8.99196975e-03, 9.83189717e-02, 3.06732170e-02, -7.63913393e-02, -4.46917377e-02, 3.47564444e-02, 5.54822870e-02, -3.05903926e-02, 2.38518044e-03, -5.33119589e-02, 5.86052053e-02, 1.50059548e-03, -2.03027464e-02, 3.82378735e-02, 3.28032067e-03, -2.67034769e-02, 3.17935310e-02, -5.46296015e-02, 1.97696351e-02, -2.22448520e-02, 4.38505001e-02, 2.44970415e-02, -1.09113343e-02, 2.73622875e-03, 8.32284316e-02, -1.77767519e-02, 4.07689884e-02, -1.45015372e-02, 7.53508434e-02, 4.51910906e-02, -8.25277641e-02, -3.41877192e-02, 1.77418068e-02, -2.11642925e-02, 7.21889213e-02, 1.94768235e-02, -2.11453009e-02, 8.44602436e-02, 2.48519853e-02, 6.11021928e-02, -1.98512268e-03, -1.13665080e-02, 2.05941293e-02, 3.30619141e-02, 2.97438656e-03, -1.60303004e-02, -4.40949984e-02, -9.18218791e-02, -5.54856025e-02, 1.34288361e-02, 5.06847687e-02, -1.10165784e-04, -5.90740182e-02, 2.76382081e-02, -7.46401250e-02, -2.08996880e-05, 9.32212397e-02, -1.14457374e-02, 2.73100566e-02, 5.75445183e-02, -5.61816094e-04, -4.73934077e-02, -4.35944721e-02, -3.52028124e-02, 1.54333077e-02, -9.50482860e-02, -2.23925672e-02, -5.45034707e-02, 3.05453204e-02, -1.12274047e-02, 3.73750143e-02, -2.16984842e-02, -1.08364470e-01, 7.26937056e-02, 1.29395060e-03, -7.43590072e-02, -3.03321686e-02, -7.87093714e-02, -1.97824650e-02, -3.12689431e-02, 1.91068836e-02, -2.56705619e-02, -5.90012893e-02, 4.39764038e-02, -3.96308303e-02, 5.48811145e-02, 2.21167058e-02, -7.17390552e-02, 1.08197704e-01, -3.33064198e-02, 1.45169953e-02, 6.05951026e-02, -2.87664887e-02, -1.29939001e-02, -9.37548559e-03, -8.96087009e-03, -1.75709426e-02, -4.91941124e-02, -7.48750120e-02, -3.38227525e-02, 1.34311384e-02, 4.94770631e-02, -2.06464231e-02, -2.02238802e-02, 6.20684102e-02, 2.73675490e-02, 3.78895476e-02, -2.29312032e-02, 2.29282468e-03, -2.35780124e-02, -3.83791104e-02, 4.52002697e-03, -1.50561752e-02, -3.13255116e-02, -1.01998383e-02, -7.87638277e-02, -3.32716666e-02, -3.95182334e-02, -9.74203497e-02, 2.18809545e-02, -7.25470260e-02, 1.37321148e-02, 2.50995047e-02, 5.10979109e-02, 4.24298123e-02, 2.75686290e-02, -4.90514887e-03, -5.27277738e-02, 7.37626553e-02, -5.29145412e-02, 3.27520235e-03, -4.49925736e-02, -2.53231432e-02, -2.29947641e-02, -1.70712546e-02, -2.77240891e-02, -3.56300585e-02, -5.66384979e-02, -6.26312867e-02, -4.54002097e-02, -1.07850032e-02, -5.15301190e-02, 5.28749228e-02, 1.83964465e-02, -2.11781319e-02, -8.78388137e-02], dtype=float32)>, 'natural_language_instruction': <tf.Tensor: shape=(), dtype=string, numpy=b'Place the can to the left of the pot.'>, 'state': <tf.Tensor: shape=(7,), dtype=float32, numpy= array([ 0.29806843, -0.11465699, 0.10782038, 0.04275148, -0.14888743, -0.31455365, 1.0001532 ], dtype=float32)>}) ('reward', <tf.Tensor: shape=(), dtype=float32, numpy=0.0>)
相关解释：
-
'action' 这一步机器人实际执行的动作（命令）。
- open_gripper: 是否打开夹爪（bool）。
- rotation_delta: 旋转增量（3 维，roll/pitch/yaw）。
- world_vector: 世界坐标系下的移动向量（3 维，x/y/z）。
- terminate_episode: 是否结束当前 episode。 作用：这是模型要学习的"输出"，即"看到当前画面后，应该执行什么动作"。
-
'is_first', 'is_last', 'is_terminal'
- is_first=True：这是当前 episode 的第一步（初始状态）。
- is_last / is_terminal：标记 episode 是否结束或达到终止状态。 作用：帮助代码知道轨迹的开始和结束边界。
-
'observation'（最重要的一部分） 机器人"看到"的当前环境信息（输入）：
- image：主相机拍摄的 RGB 图像（480×640×3）。这是原始像素数据，和 LeWorldModel 里从 pixels 学习的输入一样。
- natural_language_embedding：自然语言指令的向量嵌入（512 维）。对应指令是 "Place the can to the left of the pot."（把罐子放到锅的左边）。
- natural_language_instruction：原始文本指令。
- state：机器人本体状态（7 维），通常包括关节角度、末端执行器位置、夹爪状态等 proprioceptive（本体感觉）信息。
-
'reward' 当前步的奖励值。这里是 0.0（大多数演示数据里 reward 都是 0 或稀疏，只有成功时才为 1）。 Open X-Embodiment 大部分是 reward-free 或稀疏奖励的演示数据（Imitation Learning 为主）。
python
import tensorflow as tf
import tensorflow_datasets as tfds
# load raw dataset --> replace this with tfds.load() on your
# local machine!
dataset = 'kuka'
b = tfds.builder_from_directory(builder_dir=dataset2path(dataset))
ds = b.as_dataset(split='train[:10]')
def episode2steps(episode):
return episode['steps']
def step_map_fn(step):
return {
'observation': {
'image': tf.image.resize(step['observation']['image'], (128, 128)),
},
'action': tf.concat([
step['action']['world_vector'],
step['action']['rotation_delta'],
step['action']['gripper_closedness_action'],
], axis=-1)
}
# convert RLDS episode dataset to individual steps & reformat
ds = ds.map(
episode2steps, num_parallel_calls=tf.data.AUTOTUNE).flat_map(lambda x: x)
ds = ds.map(step_map_fn, num_parallel_calls=tf.data.AUTOTUNE)
# shuffle, repeat, pre-fetch, batch
ds = ds.cache() # optionally keep full dataset in memory
ds = ds.shuffle(100) # set shuffle buffer size
ds = ds.repeat() # ensure that data never runs outDROID（Distributed Robot Interaction Dataset）
- 76k+ 演示轨迹（约 350 小时），覆盖 564 个真实场景、86 个任务，由 50 人收集。in-the-wild（真实家庭/办公室环境），多视角相机。
- 著名项目：用于训练通用操纵策略，许多后续真实机器人论文引用，作为 in-the-wild 基准。
python
# Full DROID dataset in RLDS (1.7TB)
gsutil -m cp -r gs://gresearch/robotics/droid <path_to_your_target_dir>
# Example 100 episodes from the DROID dataset in RLDS for debugging (2GB)
gsutil -m cp -r gs://gresearch/robotics/droid_100 <path_to_your_target_dir>
# Raw DROID dataset in stereo HD, stored as MP4 videos (8.7TB)
gsutil -m cp -r gs://gresearch/robotics/droid_raw <path_to_your_target_dir>
# Raw DROID dataset, non-stereo HD video only (5.6TB, excluding stereo video & raw SVO cam files)
gsutil -m rsync -r -x ".*SVO.*|.*stereo.*\.mp4$" "gs://gresearch/robotics/droid_raw" <path_to_your_target_dir>数据集的 Schema：（单条episode）
pythonDROID = { "episode_metadata": { "recording_folderpath": tf.Text, # path to the folder of recordings "file_path": tf.Text, # path to the original data file }, "steps": { "is_first": tf.Scalar(dtype=bool), # true on first step of the episode "is_last": tf.Scalar(dtype=bool), # true on last step of the episode "is_terminal": tf.Scalar(dtype=bool), # true on last step of the episode if it is a terminal step, True for demos "language_instruction": tf.Text, # language instruction "language_instruction_2": tf.Text, # alternative language instruction "language_instruction_3": tf.Text, # alternative language instruction "observation": { "gripper_position": tf.Tensor(1, dtype=float64), # gripper position state "cartesian_position": tf.Tensor(6, dtype=float64), # robot Cartesian state "joint_position": tf.Tensor(7, dtype=float64), # joint position state "wrist_image_left": tf.Image(180, 320, 3, dtype=uint8), # wrist camera RGB left viewpoint "exterior_image_1_left": tf.Image(180, 320, 3, dtype=uint8), # exterior camera 1 left viewpoint "exterior_image_2_left": tf.Image(180, 320, 3, dtype=uint8), # exterior camera 2 left viewpoint }, "action_dict": { "gripper_position": tf.Tensor(1, dtype=float64), # commanded gripper position "gripper_velocity": tf.Tensor(1, dtype=float64), # commanded gripper velocity "cartesian_position": tf.Tensor(6, dtype=float64), # commanded Cartesian position "cartesian_velocity": tf.Tensor(6, dtype=float64), # commanded Cartesian velocity "joint_position": tf.Tensor(7, dtype=float64), # commanded joint position "joint_velocity": tf.Tensor(7, dtype=float64), # commanded joint velocity }, "discount": tf.Scalar(dtype=float32), # discount if provided, default to 1 "reward": tf.Scalar(dtype=float32), # reward if provided, 1 on final step for demos "action": tf.Tensor(7, dtype=float64), # robot action, consists of [6x joint velocities, 1x gripper position] }, }episode_metadata（元数据）
- recording_folderpath：原始录制文件夹路径
- file_path：原始数据文件路径
steps（每一步 / 每一帧）
每个 episode 由很多 steps 组成，每一个 step 包含：
- is_first：是否是这条轨迹的第一帧
- is_last：是否是最后一条
- is_terminal：是否是任务结束
- 语言指令 （非常重要）：
- language_instruction：主要指令（英文）
- language_instruction_2 / 3：备选指令（同一个任务可能有不同说法）
- observation（观测 / 输入） ：
- gripper_position：夹爪位置
- cartesian_position：机器人末端执行器在空间中的位置（6维）
- joint_position：7个关节的角度
- 图像 （3个相机视角，左视图）：
- wrist_image_left：手腕相机图像（180×320）
- exterior_image_1_left：外部相机1
- exterior_image_2_left：外部相机2
- action_dict（动作指令） ：
- 包含 gripper、cartesian、joint 的位置和速度命令
- action （最终使用的动作）：
- 7维向量 = 6个关节速度 + 1个夹爪位置
- reward：奖励值（演示数据中通常最后一步是 1，其余是 0）
- discount：折扣因子（一般是 1）
BridgeData V2
- 是什么：大规模机器人操纵演示数据集（大量 pick-and-place 等任务），有 teleoperated（人工遥操）和 scripted 数据。
- 用途：大规模机器人学习、策略训练、数据增强。
- 下载 ：BridgeData V2 （JPEGS + TFDS 格式）
- 著名项目：Berkeley 的 Octo 模型等，许多 scalable robot learning 工作。
ALOHA（特别是 Mobile ALOHA）
- 是什么：低成本双臂遥操硬件 + 数据集，专注于双臂灵巧操纵（bimanual manipulation），包括移动底座版本。
- 用途：双臂任务、灵巧操作、变形物体处理。
- 下载 ：项目页面 Mobile ALOHA 或 Hugging Face（搜索 ALOHA 数据集），GitHub https://github.com/tonyzhaozh/aloha
- 著名项目：Mobile ALOHA 项目，许多扩散策略（Diffusion Policy）论文使用。
LIBERO
- 是什么：Lifelong Robot Learning 基准，包含多个任务套件（Spatial、Object、Goal 等），高质量演示。
- 用途：终身学习、知识迁移、基准测试。
- 下载 ：LIBERO Datasets -- LIBERO （Box 链接）或 Hugging Face（HuggingFaceVLA/libero 或 physical-intelligence/libero）
- 著名项目：终身学习、Pi0 等模型基准。
Code Structures
pythonlibero/ assets/ # Where all the object mesh files are stored bddl_files/ # PDDL file definitions for tasks libero_goal/*.bddl # 10 tasks of LIBERO-Goal suite libero_spatial/*.bddl # 10 tasks of LIBERO-Spatial suite libero_object/*.bddl # 10 tasks of LIBERO-Object suite libero_10/*.bddl # 10 tasks of LIBERO-100 for evaluation (aka LIBERO-LONG) libero_90/*.bddl # 90 tasks of LIBERO-100 for pretraining benchmark/ # Task orders for evaluation of all benchmarks envs/ # Environment definitions for LIBERO tasks init_files/ # Fixed initializations for benchmark evaluation utils/ # Miscellaneous utility functions关于task的部分：
pythonfrom libero.libero import benchmark from libero.libero.envs import OffScreenRenderEnv benchmark_dict = benchmark.get_benchmark_dict() task_suite_name = "libero_10" # can also choose libero_spatial, libero_object, etc. task_suite = benchmark_dict[task_suite_name]() # retrieve a specific task task_id = 0 task = task_suite.get_task(task_id) task_name = task.name task_description = task.language task_bddl_file = os.path.join(get_libero_path("bddl_files"), task.problem_folder, task.bddl_file) print(f"[info] retrieving task {task_id} from suite {task_suite_name}, the " + \ f"language instruction is {task_description}, and the bddl file is {task_bddl_file}") # step over the environment env_args = { "bddl_file_name": task_bddl_file, "camera_heights": 128, "camera_widths": 128 } env = OffScreenRenderEnv(**env_args) env.seed(0) env.reset() init_states = task_suite.get_task_init_states(task_id) # for benchmarking purpose, we fix the a set of initial states init_state_id = 0 env.set_init_state(init_states[init_state_id]) dummy_action = [0.] * 7 for step in range(10): obs, reward, done, info = env.step(dummy_action) env.close()内容物：
找到文件，真实路径在: /content/LIBERO/libero/libero/bddl_files/libero_10/LIVING_ROOM_SCENE2_put_both_the_alphabet_soup_and_the_tomato_sauce_in_the_basket.bddl --- 以下是 BDDL 文件内容 --- (define (problem LIBERO_Living_Room_Tabletop_Manipulation) (:domain robosuite) (:language put both the alphabet soup and the tomato sauce in the basket) (:regions (basket_init_region (:target living_room_table) (:ranges ( (-0.01 0.25 0.01 0.27) ) ) (:yaw_rotation ( (0.0 0.0) ) ) ) (milk_init_region (:target living_room_table) (:ranges ( (0.025 -0.125 0.07500000000000001 -0.07500000000000001) ) ) (:yaw_rotation ( (0.0 0.0) ) ) ) (cream_cheese_init_region (:target living_room_table) (:ranges ( (0.07500000000000001 -0.225 0.125 -0.17500000000000002) ) ) (:yaw_rotation ( (0.0 0.0) ) ) ) (orange_juice_init_region (:target living_room_table) (:ranges ( (-0.025 -0.275 0.025 -0.225) ) ) (:yaw_rotation ( (0.0 0.0) ) ) ) (tomato_sauce_init_region (:target living_room_table) (:ranges ( (-0.125 0.025 -0.07500000000000001 0.07500000000000001) ) ) (:yaw_rotation ( (0.0 0.0) ) ) ) (alphabet_soup_init_region (:target living_room_table) (:ranges ( (-0.125 -0.175 -0.07500000000000001 -0.125) ) ) (:yaw_rotation ( (0.0 0.0) ) ) ) (butter_init_region (:target living_room_table) (:ranges ( (0.025 0.025 0.07500000000000001 0.07500000000000001) ) ) (:yaw_rotation ( (0.0 0.0) ) ) ) (ketchup_init_region (:target living_room_table) (:ranges ( (-0.275 -0.175 -0.225 -0.125) ) ) (:yaw_rotation ( (0.0 0.0) ) ) ) (contain_region (:target basket_1) ) ) (:fixtures living_room_table - living_room_table ) (:objects alphabet_soup_1 - alphabet_soup cream_cheese_1 - cream_cheese tomato_sauce_1 - tomato_sauce ketchup_1 - ketchup orange_juice_1 - orange_juice milk_1 - milk butter_1 - butter basket_1 - basket ) (:obj_of_interest alphabet_soup_1 tomato_sauce_1 basket_1 ) (:init (On alphabet_soup_1 living_room_table_alphabet_soup_init_region) (On cream_cheese_1 living_room_table_cream_cheese_init_region) (On tomato_sauce_1 living_room_table_tomato_sauce_init_region) (On ketchup_1 living_room_table_ketchup_init_region) (On milk_1 living_room_table_milk_init_region) (On orange_juice_1 living_room_table_orange_juice_init_region) (On butter_1 living_room_table_butter_init_region) (On basket_1 living_room_table_basket_init_region) ) (:goal (And (In alphabet_soup_1 basket_1_contain_region) (In tomato_sauce_1 basket_1_contain_region)) ) )task解析：
把 alphabet soup（字母汤罐）和 tomato sauce（番茄酱）都放到篮子里
文件逐段详细解释
lisp
(define (problem LIBERO_Living_Room_Tabletop_Manipulation) (:domain robosuite) (:language put both the alphabet soup and the tomato sauce in the basket)
- problem：这是任务的名字。
- domain：基于 robosuite（MuJoCo 的机器人仿真框架）。
- :language ：自然语言指令 ------ 这就是模型或人类看到的任务描述。
(:regions ...) ------ 区域定义
这一段定义了桌子上不同物品的初始放置区域（范围）：
- basket_init_region：篮子应该放在桌子的哪个位置。
- alphabet_soup_init_region、tomato_sauce_init_region 等：每种物品的初始放置范围（x, y 坐标区间）。
- contain_region：篮子里面的"可容纳区域"。
这些区域保证了任务的可重复性（每次初始化位置都在一定范围内）。
(:fixtures ...)
lisp
(:fixtures living_room_table - living_room_table )定义了场景中的固定物体（这里只有一张 living_room_table）。
(:objects ...)
lisp
(:objects alphabet_soup_1 - alphabet_soup tomato_sauce_1 - tomato_sauce ... basket_1 - basket )列出这个场景中所有出现的物体，并指定它们的类型。
(:obj_of_interest ...)
lisp
(:obj_of_interest alphabet_soup_1 tomato_sauce_1 basket_1 )关键物体 ：只有这三个物体是完成任务真正需要的，其他物体（如牛奶、橙汁等）是干扰物，增加任务难度。
(:init ...) ------ 初始状态
lisp
(:init (On alphabet_soup_1 living_room_table_alphabet_soup_init_region) (On tomato_sauce_1 living_room_table_tomato_sauce_init_region) ... )定义任务开始时每个物体放在哪里。
(:goal ...) ------ 目标条件（最重要！）
lisp
(:goal (And (In alphabet_soup_1 basket_1_contain_region) (In tomato_sauce_1 basket_1_contain_region)) )成功条件 ： 字母汤罐 和 番茄酱 都必须放在篮子里面。
只有当这个条件满足时，任务才算完成，reward 才会给 +1。
LeRobotDataset
- 是什么 ：Hugging Face 推出的标准化机器人数据集格式（v3.0），支持 streaming、多模态、易可视化。不是单一数据集，而是许多数据集的统一入口。
- 用途：方便加载/处理各种机器人数据，与 LeRobot 库无缝集成。
- 下载/访问 ：https://huggingface.co/lerobot （搜索数据集）或通过 LeRobot 库加载。
- 著名项目：Hugging Face LeRobot 框架，许多开源机器人项目采用此格式。