# World Actions

> Generative Studio for Physical AI. World Actions turns one real-world capture into many: new scenes, lighting and embodiments, labeled frame by frame and replayed in simulation. Run video, perception, tracking, 3D and robot-policy models on your episodes and export them to LeRobot, RLDS, MCAP, ROS 2 bags or HDF5.

Action Agent: open a task (versions, variations, labels, sim replays — whatever you need), submit your videos, and the agent plus a human reviewer deliver it complete in 12–24 hours.

Pricing uses one token balance. Studio shows usage for the selected action, model and workload before a run.

## Pages

- [Home](https://actions.world/): what World Actions does — multiply reality, label it, simulate it, export it
- [Models](https://actions.world/models): all 43 integrated models with licences and open-weight availability
- [Datasets](https://actions.world/datasets): 22 robot and egocentric datasets (World Actions captures and Hugging Face LeRobot datasets) to run actions on
- [Process my videos](https://actions.world/process): run actions yourself in the Studio, or hand a task to the Action Agent (12–24 h)
- [Actions for Students](https://actions.world/students): Free for students: the Starter plan, with 3,000 tokens every month, for 6 months. Apply online.
- [Actions for Startups](https://actions.world/startups): Free for early-stage startups: 25,000 tokens to spend on any action and model. Apply online.
- [Notes](https://actions.world/notes): publications, releases and milestones

## Features

- Search Filter: Search indexed videos with natural language.
- Cleanup: Remove content not needed in the video.
- Segmentation: Box objects and extract task plus low-level action labels.
- Hand Pose: Estimate hand landmarks for finger skeleton overlay.
- Depth Map: Estimate depth maps for colorized video overlay.
- Diffusion: Generate styled video variations using state-of-the-art video-to-video models.
- Image to Simulator: Reconstruct physics-ready MuJoCo simulations from multi-view images.
- Analyze: Create data points to learn better behavior.
- Full Body Pose: Track 33 body landmarks in image and hip-centered 3D coordinates, and preview a skeleton over the video.
- Face Landmark: Track 478 face landmarks plus 52 expression scores, and preview a face mesh over the video.
- Gesture Detection: Label each hand with a gesture and confidence, and preview the name on the hand.
- Holistic Landmark: Track 543 pose, face, and hand landmarks together (33 + 468 + 42), and preview them on the video.
- Mask Segmentation: Trace each detected object’s outline as a polygon, separate from box segmentation.

## Models

- [Gemini 3.8 Flash](https://actions.world/models/gemini-3.8-flash) (Google, Analyze, Cleanup, Mask Segmentation, Segmentation): Sharper scene summaries. Balanced cleanup detection. Accurate object outlines. Tighter HUD boxes
- [Gemini 4 Argon](https://actions.world/models/gemini-4-argon) (Google, Analyze, Cleanup, Mask Segmentation, Segmentation): Frame-by-frame scene reads. Sharpest idle-footage trims. General-purpose masks. Precise boxes and labels
- [GPT-6 Astra Fast](https://actions.world/models/gpt-6-astra-fast) (OpenAI, Analyze, Cleanup, Mask Segmentation, Segmentation): Faster dataset sweep. Faster high-quality cleanup. Faster mask segmentation. Faster Astra labelling
- [GPT-6 Astra](https://actions.world/models/gpt-6-astra) (OpenAI, Analyze, Cleanup, Image to Simulator, Mask Segmentation, Segmentation): Deep sequence reasoning. Highest-accuracy cleanup. Photos to MuJoCo. Highest-quality boundaries. Highest-fidelity box labels
- [Qwen 3.8 Flash](https://actions.world/models/qwen-3-8-flash) (Alibaba, Analyze, Cleanup, Mask Segmentation, Segmentation): Cheap long-clip analysis. Cheap large-dataset cleanup. Budget bulk masks. Budget bulk annotation
- [Qwen 3.8 Max](https://actions.world/models/qwen-3-8-max) (Alibaba, Analyze, Cleanup, Image to Simulator, Mask Segmentation, Segmentation): Detailed behavior points. Dense overlapping scenes. Budget scene reconstruction. Precise dense-scene boundaries. Dense overlapping objects
- [GPT-6 Luna](https://actions.world/models/gpt-6-luna) (OpenAI, Analyze, Cleanup, Image to Simulator, Mask Segmentation, Segmentation): General Luna scene reads. General-purpose Luna cleanup. Luna photos to MuJoCo. General Luna boundaries. General Luna box labels
- [GPT-6 Luna Fast](https://actions.world/models/gpt-6-luna-fast) (OpenAI, Analyze, Cleanup, Mask Segmentation, Segmentation): Faster Luna dataset sweep. Faster Luna cleanup. Faster Luna masks. Faster Luna labelling
- [GPT-6 Luna Decision](https://actions.world/models/gpt-6-luna-decision) (OpenAI, Analyze, Cleanup, Image to Simulator, Mask Segmentation, Segmentation): Decision-led scene reads. Decision-led cleanup cuts. Decision-led scene rebuilds. Decision-led mask boundaries. Decision-led box labels
- [GPT-6 Sol](https://actions.world/models/gpt-6-sol) (OpenAI, Analyze, Cleanup, Image to Simulator, Mask Segmentation, Segmentation): Sol sequence reasoning. Sol reasoning for cleanup. Sol photos to MuJoCo. Sol reasoning for masks. Sol reasoning for labels
- [GPT-6 Sol Fast](https://actions.world/models/gpt-6-sol-fast) (OpenAI, Analyze, Cleanup, Mask Segmentation, Segmentation): Faster Sol dataset sweep. Faster Sol cleanup. Faster Sol masks. Faster Sol labelling
- [GPT-6.1 Sol](https://actions.world/models/gpt-6-1-sol) (OpenAI, Analyze, Cleanup, Image to Simulator, Mask Segmentation, Segmentation): Newer Sol scene reasoning. Newer Sol cleanup. Newer Sol scene rebuilds. Newer Sol boundaries. Newer Sol box labels
- [GPT-6.1 Sol Fast](https://actions.world/models/gpt-6-1-sol-fast) (OpenAI, Analyze, Cleanup, Mask Segmentation, Segmentation): Faster GPT-6.1 sweep. Faster GPT-6.1 cleanup. Faster GPT-6.1 masks. Faster GPT-6.1 labelling
- [Depth Anything V3](https://actions.world/models/da3-metric-large) (ByteDance, Depth Map): Metric depth overlays
- [Depth Anything V2](https://actions.world/models/depth-anything-v2) (ByteDance, Depth Map): Relative depth per frame
- [DepthCrafter](https://actions.world/models/depthcrafter) (Tencent, Depth Map): Consistent depth across the clip
- [MiDaS](https://actions.world/models/midas) (Intel, Depth Map): Relative inverse depth per frame
- [Depth Pro](https://actions.world/models/depth-pro) (Apple, Depth Map): Metric depth in meters
- [Marigold](https://actions.world/models/marigold) (ETH Zurich, Depth Map): Diffusion depth per frame
- [FoundationStereo](https://actions.world/models/foundation-stereo) (NVIDIA, Depth Map): Metric depth from a side-by-side stereo frame
- [Wan 3.0 Prime](https://actions.world/models/wan-3-prime) (Alibaba, Diffusion): Higher-quality Wan restyle
- [Wan 3.0](https://actions.world/models/wan-3) (Alibaba, Diffusion): Clip restyle from Wan
- [Gemini Omni Flash 1.1](https://actions.world/models/gemini_omni_flash_1.1) (Google, Diffusion): Shorter Omni look shifts
- [MiniMax H3](https://actions.world/models/hailuo3) (MiniMax, Diffusion): Reference restyle up to 15s
- [Seedance 2.5](https://actions.world/models/seedance2_5) (ByteDance, Diffusion): Longer restyle windows
- [Aleph 2](https://actions.world/models/aleph2) (Runway, Diffusion): Prompted environment swaps
- [Gemini Omni Flash](https://actions.world/models/gemini_omni_flash) (Google, Diffusion): Weather and lighting shifts
- [Seedance 2.0 Fast](https://actions.world/models/seedance2_fast) (ByteDance, Diffusion): Cheap look iteration
- [Seedance 2.0 Mini](https://actions.world/models/seedance2_mini) (ByteDance, Diffusion): Bulk short-clip restyle
- [Grok Imagine Video](https://actions.world/models/grok-imagine-video) (xAI, Diffusion): Prompted edit of the clip
- [Seedance 2.0](https://actions.world/models/seedance2) (ByteDance, Diffusion): Locked-camera video restyle
- [MediaPipe Face Landmarker](https://actions.world/models/mediapipe-face-landmarker) (Google, Face Landmark): 478 face landmarks
- [MediaPipe Pose Landmarker Heavy](https://actions.world/models/mediapipe-pose-landmarker-heavy) (Google, Full Body Pose): Highest-quality body tracking
- [MediaPipe Pose Landmarker Full](https://actions.world/models/mediapipe-pose-landmarker-full) (Google, Full Body Pose): Faster body tracking
- [MediaPipe Pose Landmarker Lite](https://actions.world/models/mediapipe-pose-landmarker-lite) (Google, Full Body Pose): Fastest body tracking
- [RF-DETR Keypoint Preview](https://actions.world/models/rfdetr-keypoint-preview) (Roboflow, Full Body Pose): 17 COCO person keypoints
- [MediaPipe Gesture Recognizer](https://actions.world/models/mediapipe-gesture-recognizer) (Google, Gesture Detection): Gesture label per hand
- [MediaPipe Holistic Landmarker](https://actions.world/models/mediapipe-holistic-landmarker) (Google, Holistic Landmark): Pose, face, and hands
- [Claude Fable 5](https://actions.world/models/claude-fable-5) (Anthropic, Image to Simulator): Primary scene reconstruction
- [Claude Opus 5](https://actions.world/models/claude-opus-5) (Anthropic, Image to Simulator): Complex environment rebuilds
- [Claude Sonnet 5](https://actions.world/models/claude-sonnet-5) (Anthropic, Image to Simulator): Balanced scene reconstruction
- [MediaPipe Hand Landmarker](https://actions.world/models/mediapipe-hand-landmarker) (Google, Hand Pose): On-device hand landmarks
- [TwelveLabs](https://actions.world/models/twelvelabs) (TwelveLabs, Search Filter): Natural-language clip search

## Datasets

- [Mug pick & place](https://actions.world/datasets/mug-pick) (World Actions, Egocentric video): One real first-person capture: a hand picks up a mug and places it in a basket. The source behind every reality on this site.
- [Humanoid mug pick](https://actions.world/datasets/humanoid-mug) (World Actions, Humanoid): The same pick re-embodied on a humanoid hand — the demonstration a humanoid policy trains on.
- [Kitchen scene → MuJoCo](https://actions.world/datasets/mujoco-kitchen) (World Actions, Simulation): A real kitchen turned into a MuJoCo scene with Astra, ready for physics replays and policy evaluation.
- [EgoDex](https://actions.world/datasets/egodex) (griffinlabs/EgoDex-LeRobot-v3.0, Egocentric video): Apple's head-mounted video of people doing dexterous tabletop tasks with their hands, with tracked hand poses — converted to LeRobot.
- [EPIC-KITCHENS-100](https://actions.world/datasets/epic-kitchens-100) (lightly-ai/epic-kitchens-100-clips, Egocentric video): First-person kitchen video, cut into one short clip per narrated action — opening the fridge, cutting an onion, washing a pan.
- [EgoExoLearn](https://actions.world/datasets/egoexolearn) (hyf015/EgoExoLearn, Egocentric video): People film first-person video while following a demonstration video of a procedural task, at home and in the lab.
- [DROID (100 episodes)](https://actions.world/datasets/droid-100) (lerobot/droid_100, Teleoperation): A 100-episode slice of DROID: a Franka arm doing everyday tasks in real homes and labs, with two outside views and a wrist camera.
- [TACO Play](https://actions.world/datasets/taco-play) (lerobot/taco_play, Teleoperation): Language-labelled play on a tabletop — rotating and stacking blocks, pressing buttons, pulling handles — from a static and a gripper camera.
- [Berkeley AUTOLab UR5](https://actions.world/datasets/berkeley-ur5) (lerobot/berkeley_autolab_ur5, Teleoperation): A UR5 arm sweeping a cloth, moving a bottle, picking up a cup and taking a toy tiger out of a bowl — with RGB, depth and a wrist camera.
- [SO-101 pick & place](https://actions.world/datasets/so101-pickplace) (lerobot/svla_so101_pickplace, Teleoperation): A low-cost SO-101 arm putting a pink LEGO brick into a transparent box, filmed from above and the side.
- [xArm lift](https://actions.world/datasets/xarm-lift) (lerobot/xarm_lift_medium, Simulation): A simulated xArm picks up a cube and lifts it — short episodes for training and testing policies quickly.
- [ALOHA cups open](https://actions.world/datasets/aloha-cups) (lerobot/aloha_static_cups_open, Bimanual): Two arms pick up a plastic cup and open its lid, seen from four cameras: high, low and both wrists.
- [Unitree H1 warehouse](https://actions.world/datasets/unitree-h1-warehouse) (lerobot/unitreeh1_warehouse, Humanoid): A Unitree H1 humanoid grabs spray paint from a shelf and drops it in a bin on a robot dog, seen from its two head cameras.
- [ALOHA sim insertion](https://actions.world/datasets/aloha-sim-insertion) (lerobot/aloha_sim_insertion_human, Simulation): Human-teleoperated peg insertion with two arms, in simulation — a standard benchmark for bimanual policies.
- [LIBERO-10](https://actions.world/datasets/libero-10) (lerobot/libero_10_image, Simulation): Ten long-horizon household tasks for a Panda arm in simulation — mugs on plates, the microwave, a moka pot — from the LIBERO benchmark.
- [Push-T](https://actions.world/datasets/pusht) (lerobot/pusht, Simulation): Push a T-shaped block into place — the classic 2D benchmark from Diffusion Policy.
- [GR00T X-Embodiment Sim](https://actions.world/datasets/groot-x-embodiment-sim) (nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim, Simulation): NVIDIA's simulated manipulation data for post-training GR00T N1 — bimanual Panda arms, GR1 humanoid tabletop tasks and kitchen tasks.
- [RoboCasa365](https://actions.world/datasets/robocasa365-pretrain) (ember-lab-berkeley/robocasa365-pretrain-mg, Simulation): Machine-generated demos of a mobile Panda doing hundreds of everyday kitchen tasks in RoboCasa simulated homes.
- [Meta-World MT50](https://actions.world/datasets/metaworld-mt50) (lerobot/metaworld_mt50, Simulation): Demonstrations for all 50 Meta-World tasks — opening drawers, pressing buttons, pushing objects — in MuJoCo.
- [D2E desktop gameplay](https://actions.world/datasets/d2e-gameplay) (open-world-agents/D2E-480p, Games): 269 hours of PC gameplay from 29 games with synced keyboard and mouse input, recorded to pretrain vision-action models for robots.
- [GameWAM ViZDoom](https://actions.world/datasets/gamewam-vizdoom) (Yunncheng/gamewam-vizdoom, Games): First-person Doom episodes played by trained agents across four scenarios, with every action and the player state per frame.
- [Counter-Strike deathmatch](https://actions.world/datasets/csgo-deathmatch) (TeaPearce/CounterStrike_Deathmatch, Games): CS:GO deathmatch screen recordings paired with mouse and keyboard actions, from online servers plus clean expert sessions.

## Notes

- [Real, synthetic and Sim](https://actions.world/notes/one-action-layer): Robot data comes from three places that never agree on format. We bring captures, generated worlds and simulation rollouts into one structured stream of actions.
- [One capture in, endless realities out](https://actions.world/notes/multiply-reality): Why we re-generate a single recorded action into many worlds — and label every one.

## Optional

- [Full text](https://actions.world/llms-full.txt): detailed models, plans, use cases and published notes