datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Cosmos3-DROID
DROID: Distributed Robot Interaction Dataset
Dataset Summary
DROID (Distributed Robot Interaction Dataset) is a large-scale "in-the-wild" robot manipulation dataset containing 76K teleoperated demonstration trajectories — approximately 350 hours of interaction data — collected across 564 unique scenes, 86 tasks, and 52 buildings over the course of 12 months. The data was collected by 50 data collectors at 18 labs across 13 institutions in North America, Asia, and… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/Cosmos3-DROID.cosmos3-ap-openarm-wam-robot-auxonly-hi-movonly-chain-lam10-aug-v2v-4gpu-b256-8kcosmos3-i2v-survival-sdg
Cosmos3 image-to-video: subject survival on synthetic rescue scenes
26 generated clips — 13 from Cosmos3-Super and 13 from Cosmos3-Nano, conditioned on the same 13 source
frames — plus the per-frame scoring that produced them. Published as the raw material behind a negative result:
image-conditioned generation does not hold a subject in place across a clip, and the bigger model does not
fix it.
Provenance — read this first
Every clip is conditioned on a frame from… See the full description on the dataset page: https://huggingface.co/datasets/ubr-physical-ai/cosmos3-i2v-survival-sdg.cosmos3-physicsiq-evidence
Cosmos3 on Physics-IQ — submission evidence
Generated videos, scorer outputs, prompts and configs backing the Cosmos3 entries on the
Physics-IQ benchmark leaderboards,
posted as evidence for
issue #75 (Super-I2V and
Nano score discrepancies) and
issue #76 (Edge submission).
Every package holds the unmodified model outputs (121 frames), the scorer-ready staged
clips (120 frames, exactly 5.000 s), the official scorer's outputs, and the exact prompts
and configs — so any number… See the full description on the dataset page: https://huggingface.co/datasets/akashgokul-nvidia/cosmos3-physicsiq-evidence.in-context-learning-cosmos3-output
Physical-ICL × Cosmos3 — generated outputs
Video-generation outputs from NVIDIA Cosmos3-Nano (Diffusers Cosmos3OmniPipeline,
image-to-video) on the Physical-ICL dataset (Vincwng/Physical-ICL, subset
physiq_prelim, 66 query samples). This studies physical in-context learning: does
showing a demonstration change how the model continues a query scene?
Total generated: 247 videos across 66 query tasks, in 6 configurations.
Configurations
Every configuration uses the… See the full description on the dataset page: https://huggingface.co/datasets/yqi19/in-context-learning-cosmos3-output.cosmos3-dpo-episode-weibull-domain30
cosmos3-dpo-episode-weibull-domain30
Open research dataset (ICRA paper DPO training pipeline for the Cosmos3 WAM joint video+action policy model, G1 humanoid + BrainCo dexterous hand robot).
Current extraction release
All episodes in this release were refreshed on 2026-09-14 with the RoboHaMeR v13 hand-pose checkpoint and a 4× video upscale before YOLO hand detection and pose inference. The raw generated videos are preserved; qpos_robohamer.npz and the derived a_!… See the full description on the dataset page: https://huggingface.co/datasets/SergeyKurchev/cosmos3-dpo-episode-weibull-domain30.cosmos3-ap-openarm-wam-robot-aug-v2v-b256-4kcosmos3-ap-openarm-wam-robot-aug-v2v-b256-8kCosmos3-DROID-Nominal-1000cosmos3_testcosmos3-physicsiq-evidence-v2vrollout_cosmos3_so101_drift_d6231_step100000_large_green_cube_sync_0810_1550_20260810_155045This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"fps": 30,
"features": {
"action": {
"dtype": "float32",
"names": [
"shoulder_pan.pos",
"shoulder_lift.pos",
"elbow_flex.pos",
"wrist_flex.pos",
"wrist_roll.pos",
"gripper.pos"
],
"shape": [
6… See the full description on the dataset page: https://huggingface.co/datasets/geonmin-kim/rollout_cosmos3_so101_drift_d6231_step100000_large_green_cube_sync_0810_1550_20260810_155045.cosmos3-ap-openarm-wam-robot-ours-aug-v2v-b256-4krollout_cosmos3_so101_drift_d6231_step100000_large_green_cube_sync_0810_1535_20260810_153608This dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"fps": 30,
"features": {
"action": {
"dtype": "float32",
"names": [
"shoulder_pan.pos",
"shoulder_lift.pos",
"elbow_flex.pos",
"wrist_flex.pos",
"wrist_roll.pos",
"gripper.pos"
],
"shape": [
6… See the full description on the dataset page: https://huggingface.co/datasets/geonmin-kim/rollout_cosmos3_so101_drift_d6231_step100000_large_green_cube_sync_0810_1535_20260810_153608.cosmos3-droid-mujocoThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "panda",
"total_episodes": 1,
"total_frames": 144,
"total_tasks": 3,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 15,
"splits": {
"train": "0:1"
},
"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/cagataydev/cosmos3-droid-mujoco.cosmos3-ap-openarm-wam-robot-ours-aug-v2v-b256-8kcosmos3-droid-mujoco-3kThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v3.0",
"robot_type": "panda",
"total_episodes": 1,
"total_frames": 3000,
"total_tasks": 6,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"fps": 15,
"splits": {
"train": "0:1"
},
"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/cagataydev/cosmos3-droid-mujoco-3k.cosmos3-t2v-release
Cosmos3 Text-to-Video Release Dataset
This is the Cosmos3 text-to-video release dataset.
Layout
videos/ contains per-config single videos for 5 configs: A_bf16, B_cache, C_cache_m1, D_cache_m1_m2, and E_cache_m1_m2_a2. Each config has 100 prompts, with files prompt_001.mp4 through prompt_100.mp4. Each video is 1280x720, 189 frames, 24fps.
compare/ contains LEFT|RIGHT side-by-side comparison groups at 2560x720. Each group has 100 prompts, with files prompt_001.mp4… See the full description on the dataset page: https://huggingface.co/datasets/oliveryanzuolu/cosmos3-t2v-release.cosmos3-ap-openarm-wam-robot-auxonly-hi-movonly-chain-lam10-aug-v2v-4gpu-b256-4kcosmos3_generated_videoscosmos3-assetsworldarena-submit-cosmos3-nano-v1cosmos3-nano-evalworldarena2-track1-submit-cosmos3-super-v1cosmos3-edge-rbench-20step