Action is scarce on the open web.
A clip shows what changed. It usually does not record the exact input, execution delay, environment state, or result behind that change.
World-model training data
Training data forworld models that act.
ActionTrajectories builds action-conditioned episodes that connect what an agent observed, the action it took, and what happened next across gameplay, simulation, and robotics.
Commercial supply
Original, licensed, and partner-supplied programs.
Scope the training target first. ActionTrajectories then identifies the suitable supply route, required signals, coverage, quality evidence, rights basis, and delivery contract.
ActionTrajectories has an affiliate relationship with Yield Guild Games and can sell YGG-supplied datasets. Exact inventory and permitted uses are confirmed in writing.
The market shift
Models can predict pixels. Acting requires consequences.
Text and passive video teach appearance and sequence. They rarely expose the control signal that caused the next state. World models need that action-to-consequence record before prediction becomes reliably controllable.
Gameplay makes cause and effect instrumentable.
Games and simulations expose repeatable tasks, dense human behavior, precise controls, event boundaries, and ground-truth state.
The sellable unit is an episode, not footage.
Observations, actions, state, outcomes, timing, lineage, rights, and quality evidence ship as one inspectable training record.
Direct answer
What is an action trajectory?
An action trajectory is a time-ordered record connecting what an agent observed, what it did, how the environment changed, and what happened as a result.
That is the category: data infrastructure for cause and effect, not another warehouse of disconnected clips.
Read the technical explanationWhere trajectories apply
World models are the category. Robotics is one application.
The same observation-action-outcome structure can teach dynamics in a game, a simulator, or an embodied system. The environment changes, but the causal training record remains inspectable.
Dense actions and ground-truth state.
Instrumented games expose controls, timing, events, outcomes, and repeatable tasks for controllable prediction.
Coverage before deployment.
Simulators make rare states, counterfactual actions, and privileged environment variables measurable at scale.
World models for physical consequences.
Robot observations and controls become useful world-model data when they are aligned with executed motion, state change, and task outcome.
Before the action. The action. What happened next.
An action trajectory is useful because every signal shares one clock: what was true before, what the agent did, and what followed.
What was true at the start
The model needs both the agent's view and the environment state.
Observation
Video, audio, depth, language, and sensor measurements.
State
Pose, physics, objects, inventory, score, contact, and system status.
What the agent did
Keys, pointer deltas, gamepad axes, policy commands, or robot controls.
What happened next
Progress, reward, event, success, failure, or terminal condition.
Evidence before access
A catalog that states what it can prove.
First-party inventory and public research references remain visibly separate. Every listing names creator, source, license, restrictions, and verification date.
Open directoryOpen reference2026-07-24
MineRL 2019
Human Minecraft demonstrations with recorded video, actions, rewards, and simulator-aligned state for imitation and reinforcement learning research.
- Creator
- MineRL Team
- Scale
- 60M+ annotated state-action pairs reported in the paper
- License
- MIT License
- Commercial use
- Rights review required
Useful data survives inspection.
The release should answer the questions a researcher, lawyer, and data engineer will ask before the first training run.
Which timestamp aligns every observation, action, and event?
What does each field mean, in which unit and coordinate frame?
Which source and transformation produced this file?
Who authorized training, delivery, and customer use?
Which checks passed, failed, or remain unknown?
Open references, honestly labeled.
These records help teams compare schemas and research uses. They are not ActionTrajectories products.
Open reference2026-07-24
Rocket Science
Synchronized four-player Rocket League video, keyboard actions, game events, and per-frame physics state used to train the MIRA multiplayer world model.
- Creator
- General Intuition and Kyutai, in collaboration with Epic Games
- Scale
- 15,769 train matches and 62 test matches, approximately 29.2 TB
- License
- CC BY-NC-SA 4.0
- Commercial use
- Non-commercial
Open reference2026-07-24
NitroGen
Per-frame gamepad action annotations for public gameplay videos, intended for behavior cloning and action-to-video world-model research.
- Creator
- NVIDIA Corporation
- Scale
- 30K annotated videos and approximately 15B annotated frames
- License
- CC BY-NC 4.0
- Commercial use
- Non-commercial
Research library
Buyers need definitions before they need a form.
Source-backed guides translate emerging terminology into dataset requirements, evaluation checks, and rights decisions.
Browse all resourcesDyna-2's Million-Hour World-Action ModelWhat Dyna Robotics' million-hour human-video result shows about cross-embodiment transfer, world modeling, and the next robotics data bottleneck.What Data Do World Models Need?A practical guide to observations, actions, state, outcomes, time alignment, and coverage for world-model training datasets.Action-Conditioned Data vs. Passive VideoHow action labels change what a world model can learn, and when large passive-video corpora still provide value.How to Evaluate Gameplay Data for World ModelsA practical evaluation guide for gameplay datasets, from control capture and engine state to rights, replay filtering, and release evidence.
Independent brand. Named operator.
ActionTrajectories is operated by 10xme Technology Inc. Dataset claims are tied to visible evidence, review dates, and primary sources.
Bring us the training target.
We will map it to the observations, actions, state, outcomes, rights, and delivery evidence the dataset must contain.