datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
X-WAM-RoboTwin
X-WAM
Unified 4D World Action Modeling from Video Priors with Asynchronous Denoising
Dataset Summary
This is the RoboTwin 2.0 fine-tuning dataset used to train the X-WAM unified 4D World Action Model. It packages dual-arm bimanual manipulation demonstrations into a unified multi-view RGB-D video + low-dimensional state/action format, where each episode provides synchronized RGB videos, depth videos, dual-arm end-effector proprioception, actions, and a… See the full description on the dataset page: https://huggingface.co/datasets/sharinka0715/X-WAM-RoboTwin.X-WAM-RoboCasa
X-WAM
Unified 4D World Action Modeling from Video Priors with Asynchronous Denoising
Dataset Summary
This is the RoboCasa fine-tuning dataset used to train the X-WAM unified 4D World Action Model. It packages single-arm kitchen manipulation demonstrations into a unified multi-view RGB-D video + low-dimensional state/action format, where each episode provides synchronized RGB videos, depth videos, end-effector proprioception, actions, and language… See the full description on the dataset page: https://huggingface.co/datasets/sharinka0715/X-WAM-RoboCasa.WAM_Psi_Egoverse_datasetwam-landscape-and-deployment-20260907
WAM 选型地图与部署实测 · 2026-09
World-Action Model（世界-动作模型）在 2026 年 9 月的开源格局调研，六个候选在
NVIDIA RTX PRO 4000 Blackwell（sm_120） 上的真实部署实测，
以及这些模型的算子在 Ascend 910C（A3） 上的存在性与性能对照。
调研覆盖约 90 个模型与系统；部署部分是在单张 24 GB 消费级工作站卡上亲手跑出来的，
包含能力契约、逐层架构、算子剖析与性能数字。
为什么做这个
大多数 WAM 论文只报成功率，很少报能力契约（几路相机、动作是什么含义、世界模型那一半推理时还在不在），
几乎不报算子构成（移植到别的加速器要手写什么）。而跨论文的成功率数字大面积不可比。
这份材料试图补上这三块，并且把每个数字的口径写清楚。
五条主要结论
1. 「WAM 比 VLA 强」目前是假设，不是已验证结论。
唯一一篇第三方统一重跑的鲁棒性研究里，纯 VLA 的 π0.5 在最严的分布外评测上得分最高，… See the full description on the dataset page: https://huggingface.co/datasets/arrow-hf/wam-landscape-and-deployment-20260907.AHA-WAM-SO101-HIL-training-assets
AHA-WAM SO101 Plug HIL Training Assets
Reproducibility bundle for the SO101 power-adapter insertion experiments.
It contains a compact, ZIP-based representation of the paths expected by the
AHA-WAM training configuration:
the released AHA-WAM-pretrained.pt initialization checkpoint;
so101_ahawam_plug.zip (the downloader extracts only task 02 and 04);
ahawam_hil_raw.zip (raw rich-v1/v2/v3 HG-DAgger recordings);
the fixed 02+04 action/state normalization statistics;
cached T5… See the full description on the dataset page: https://huggingface.co/datasets/Jill111/AHA-WAM-SO101-HIL-training-assets.cosmos3-ap-openarm-wam-robot-auxonly-hi-movonly-chain-lam10-aug-v2v-4gpu-b256-8kwam_libero_test
FastWAM LIBERO long evaluation
200 rollout videos (20 trials for each of 10 tasks), 188 successes: 94.0%.
Download all videos as ZIP
Browse task folders
Episode index
Evaluation report and configuration
Videos retain their original encoding. Task and trial IDs are zero-based; filenames mark success or failure. The ZIP also contains the evaluation index and original report.
Self-Collected
Selfcollected Dataset
English · 中文
Real-robot data collected with SONIC for WB-WAM task post-training. LeRobot v3.0, 20 Hz, RGB 360 × 270. The release contains 1,011 episodes, 242,668 frames, 8 tasks, and 12 independent recordings (3.370 hours).
Each <task>/record_XXXX/ is an independent LeRobot dataset. Recordings of the same task remain separate:
<task>/record_XXXX/
data/chunk-000/file-000.parquet
videos/<camera>/chunk-000/file-000.mp4
meta/info.json
meta/stats.json… See the full description on the dataset page: https://huggingface.co/datasets/WB-WAM/Self-Collected.Pico
PICO Dataset
English · 中文
Pico motion-transfer data for WB-WAM intermediate training. LeRobot v3.0, 20 Hz, RGB 360 × 270. The release contains 13,396 episodes, 1,579,028 frames, 73 tasks, and 302 independent recordings (21.931 hours).
Each <task>/record_XXXX/ is an independent LeRobot dataset. Recordings of the same task remain separate:
<task>/record_XXXX/
data/chunk-000/file-000.parquet
videos/<camera>/chunk-000/file-000.mp4
meta/info.json
meta/stats.json
meta/tasks.parquet… See the full description on the dataset page: https://huggingface.co/datasets/WB-WAM/Pico.worldarena-track1-hz-wamwam_realdatasetopenarm-wam-v1-robot
openarm_wam_v1
Self-contained OpenArm training root for the WAM runs. Created 2026-09-15 by copying the
11 subsets actually used in training out of /data/huiwon/data/openarm_dataset_192x256/
(source left untouched) and stripping the human instruction prefix.
Layout is identical to the source (robot/..., human_as_openarm28/...), so
gr00t/configs/data/openarm_all_rel_ah24_config.py and the yamls only need the root path
changed. No symlinks: the videos/ dirs of the four… See the full description on the dataset page: https://huggingface.co/datasets/taeyoungrlwlrd/openarm-wam-v1-robot.openarm-wam-v1-cotrain
openarm_wam_v1
Self-contained OpenArm training root for the WAM runs. Created 2026-09-15 by copying the
11 subsets actually used in training out of /data/huiwon/data/openarm_dataset_192x256/
(source left untouched) and stripping the human instruction prefix.
Layout is identical to the source (robot/..., human_as_openarm28/...), so
gr00t/configs/data/openarm_all_rel_ah24_config.py and the yamls only need the root path
changed. No symlinks: the videos/ dirs of the four… See the full description on the dataset page: https://huggingface.co/datasets/taeyoungrlwlrd/openarm-wam-v1-cotrain.cosmos3-ap-openarm-wam-robot-ours-aug-v2v-b256-4kcosmos3-ap-openarm-wam-robot-aug-v2v-b256-4kfranka_WAMs_end_effectorThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"fps": 30,
"features": {
"observation.images.image": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channel"
],
"info": {
"video.height":… See the full description on the dataset page: https://huggingface.co/datasets/Grigorij/franka_WAMs_end_effector.Franka_WAMsThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"fps": 30,
"features": {
"observation.images.third_person": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channel"
],
"info": {… See the full description on the dataset page: https://huggingface.co/datasets/Grigorij/Franka_WAMs.cosmos3-ap-openarm-wam-robot-ours-aug-v2v-b256-8kcosmos3-ap-openarm-wam-robot-auxonly-hi-movonly-chain-lam10-aug-v2v-4gpu-b256-8k_movingfranka_WAMs_end_effector_fixedThis dataset was created using LeRobot.
Dataset Description
Repaired copy of Grigorij/franka_WAMs_end_effector.
The original is left untouched.
What was wrong. The recording loop initialised the previous commanded quaternion
before the task had set its commanded orientation. As a result the first frame of
every episode carried a bogus rotation delta: action[3:6] encoded the absolute
approach orientation instead of a near-zero delta. Exactly 40 frames (one per episode,
all… See the full description on the dataset page: https://huggingface.co/datasets/Grigorij/franka_WAMs_end_effector_fixed.cosmos3-ap-openarm-wam-robot-auxonly-hi-movonly-chain-lam10-aug-v2v-4gpu-b256-4kcosmos3-ap-openarm-wam-robot-aug-v2v-b256-8k_movingopenarm_wam_v1_robotWAM_Benchmark_20260728_161610This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"fps": 30,
"features": {
"action": {
"dtype": "float32",
"shape": [
6
],
"names": [
"shoulder_pan.pos",
"shoulder_lift.pos",
"elbow_flex.pos",
"wrist_flex.pos",
"wrist_roll.pos",
"gripper.pos"… See the full description on the dataset page: https://huggingface.co/datasets/AnonymousMouse404/WAM_Benchmark_20260728_161610.so100_testThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.0",
"robot_type": "so100",
"total_episodes": 2,
"total_frames": 755,
"total_tasks": 1,
"total_videos": 4,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 30,
"splits": {
"train": "0:2"
},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/wampum/so100_test.Long-WAM-YAM-Real-Eval-Rollouts
Long-WAM real-robot evaluation rollouts (YAM)
Closed-loop evaluation episodes of the Long-WAM policy on the bimanual YAM station, plus phone-camera recordings.
Mirror of the GEAR S3 bucket (2026-09-11).
Layout
Results
task
instruction
episodes
success
success rate
mean steps
mean duration
put_bricks
Put bricks into their color-matched bowls.
20
16
80%
1188
84.5 s
put_dumplings
Put all dumplings into the pan.
20
16
80%
1231
87.8 s… See the full description on the dataset page: https://huggingface.co/datasets/AaronHuangWei/Long-WAM-YAM-Real-Eval-Rollouts.wam_0428This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "so_follower",
"total_episodes": 12,
"total_frames": 22568,
"total_tasks": 1,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 30,
"splits": {
"train": "0:12"
},
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/chen4803/wam_0428.wambo-drive-cap-01