Every episode, scored
A clean pass earns full reward, partial work earns partial credit, and a faked scene earns nothing. The verdict comes from the saved file, reopened where the agent can't reach it.
Once, sailors reaching a new shore held their breath, face to face with something as large as their capacity for wonder.
We think it is happening again. We are building the next new world, and the minds that will explore it with us.
Data from the tools where 3D work actually happens.
Game engines, Blender and 3D production. For each one we provide working projects, recorded expert sessions and gradeable tasks. Use our standard datasets or have one built to your spec.
01 / 03
Game worlds
Worlds that run, and the people who know how to play them.
- aGame repositoriesComplete Unity (C#) and Unreal Engine 5 (C++) projects that compile and run, with the assets they need.
- bGameplay capture & action traceScreen, keyboard and mouse, timeline and task events, captured together: what was seen, what was done, what happened next.
- cMultimodal game stateFrames, depth, object state, events and camera parameters, drawn from a running engine and aligned to every action.
- dAgent tasks & RL environmentsTasks built on running games and their toolchains, from play and strategy to editing, engineering and testing, each with a starting state, actions, feedback, an end condition and a reward or verifier.
02 / 03
Blender & 3D tools
The craft of making, written down one step at a time.
- aCode & agent tracesTasks completed through Blender's Python API: the scripts, and the full path of observation, call, feedback and correction that led to the final .blend.
- bGUI & computer-use tracesThe same craft through the interface: screen recordings, step captures, mouse and keyboard, timestamps and step-level reasoning. Mixed GUI-and-code work is labelled action by action.
- cTasks & RL environmentsBlender workflows as environments: modeling, layout, materials, animation, camera and rendering, inspection and repair, each with a task definition, a reward or verifier, and a containerised runtime.
03 / 03
Space, form & motion
The shape of the physical world, in a form machines can read.
- a3D assetsCharacters, figures, props, vehicles, furniture, industrial parts and CAD, organised by geometry, topology, material and use.
- b3D scenesWhole spaces, not single objects: editable scene files, materials, instance and semantic labels, camera paths and walkthrough video.
- cAnimationPeople, animals, machines and robots, moving objects and multi-character interaction, with editable rigs, keyframes and actions.
- dVideo-to-3D reconstructionReference video rebuilt as an editable 3D blockout that keeps position, scale, orientation, occlusion and timing intact.
- eProcedural scenes & simulationScenes written in code and run as physics. A ramp, a box and its parameters in; scene code, a runnable project, the video and every position and velocity out.
- fCode-to-videoVideo made by code, not by a video model: layout, motion, camera and sound, all traceable to the instructions that made them.
Real tools. Real graders. Every step on the record.
Three views of what we deliver for agents that work in 3D, drawn from real samples. The full versions come in the sample pack.
RenderedGeometry only
Blender tasks a grader can hold an agent to.
The agent receives a brief and a starting project, works in Blender and submits the saved file. The grader reopens that file itself. Every part of the scene has to be real, editable geometry.
- Instruction and starting projectFresh inputs for every episode
- Pinned Blender runtimeContainerised, non-root, no network
- Host-side graderReopens the submission in an isolated worker
- Anchored rubricEach criterion names the evidence it needs
- Adversarial tests and validation reportShortcuts tested before delivery
- Harbor-compatible exportDrops into an existing rollout stack
01 · New project
redacted
redacted
Complete sessions, not clips.
A complete Blender workflow from the welcome screen to the saved file, recorded in one take. Each action is logged against the video, so the recording, the log and the final project can be checked against each other.
- Continuous screen recordingUncut, real speed, full window
- Ordered action logWall-clock and monotonic timestamps
- Scene state before and after each actionObjects, selection, frame, camera
- Input eventsWindow and screen coordinates
- The saved, editable project
- Frame-by-frame check of the resultRun in a separate process after saving
The session's resultPlayed back through the camera the session set up.
Graders that can't be gamed
Every environment is attacked before it ships. The shortcuts an agent might try are tested and closed, so reward means the work was done.
State on every step
Scene state before and after each action, input events with screen coordinates, and the saved project, checked frame by frame in a separate process.
Also built to order: game capture and RL · video-to-3D · procedural simulation · code-to-video
For teams training agents to work in 3D tools and games.
Model and agent teams use our data to teach and test work that is hard to collect in-house: modeling and scene building in Blender, playing and building in game engines, and long sessions in real desktop tools.
Post-training
Blender bpy code traces and GUI sessions as SFT data, with every step from the brief to the saved .blend.
RL & agent training
Containerised Blender and game tasks with host-side graders, so agents can be trained against a reward they can't fake.
Evaluation
Graded 3D and game tasks with adversarial checks, to see whether an agent really built the scene or only made it look right.
Data procurement
Datasets delivered with a sample pack, technical spec, quality report and rights memo, so legal and procurement can review them.
Researchers
Our PhD research team, from computer graphics, vision and simulation, designs every task and writes the grader that scores it.
Practitioners
Our 3D artists, Blender developers and engine programmers work in these tools every day, and we hold every task to the standard of their own professional work.
Method
We specify every task with a researcher, build it with a practitioner, attack it before it ships, and have a second specialist review it.
From a first sample to a living dataset.
- 01
Request samples
Tell us what you're teaching.
- 02
Evaluate
Your team tests the sample pack on your own models.
- 03
Scope
Taxonomy, volume and terms, agreed together.
- 04
Pilot
A first delivery, measured against acceptance criteria.
- 05
Scale
Full delivery, then new worlds as your models grow.
Show us what you're teaching. We'll show you a world to teach it in.
Tell us about the model and the capability you're after. We'll send a matched sample pack and dataset overview.
Prefer email? team@nexlin.ai
Thank you.
We'll reply within business days with a sample pack chosen for your work.