Pleiada

The human behavior
layer for AI.

We turn real human gameplay into frame-synced training data
video and raw input, consent secured at source.

Get in Touch →

CUSTOM VIRTUAL WORLDS

Pleiada Playgrounds

Custom-built virtual worlds designed to client specification, played by our real-player community.

Client-specified data capture from proprietary custom-built virtual worlds, played by our real-player community. Video + inputs + full engine-level world state.

◆

Owned IP

Zero publisher dependency, zero rights risk.

◆

Built to spec

Each world can be tailored to the lab's target behavior.

◆

Real physics, 3D

Realistic environments, single- or multiplayer.

◆

Infinitely extensible

New worlds and challenges on demand.

SAMPLE
3D Multiplayer Custom World Pleiada Labs
Video · inputs · engine-level world state
60 Hz WORLD STATE
Multi-agent PER-PLAYER STREAMS
100% GROUND TRUTH
Engine-level SIM EXPORT
WORLD STATE LIVE engine-level · 60 Hz
FRAME f 1042 · t 17.37s · sim_clock synced
AGENTt 17.37s p2 pos:(14.2, 0.0, -8.6) · vel:(1.1, 0, -0.4) · yaw:132°
tracking: 4 agents · 3 dynamic props · 1 lift platform · 1 world clock
CAMERA p2 pose:(14.9, 1.7, -9.1) · yaw:128° · pitch:-4° · fov:78°
INPUT p2 keys:W | Shift · action:"carry_crate" · mouse_dx:-2.1
EVENTt 17.94s pickup · agent:p2 · object:crate_07 · grip:two_handed
PHYSICS collision crate_07 ↔ floor · impulse:2.3 · restitution:0.1 @ t 18.10s
AGENTt 18.62s p4 pos:(21.6, 2.4, -3.2) · on:lift_platform · state:"ascending"
lift_platform height:2.4m → 3.1m · load:1 agent · actuator:nominal
INPUT p1 keys:A · action:"reposition" · p3 idle:0.8s
STATE objective:"stack_crates" · progress:3/5 · elapsed:174s
EVENTt 19.41s drop · agent:p2 · object:crate_07 · target:stack_zone_A
PHYSICS stack stability:0.94 · contacts:6 · resting @ t 19.6s
per-agent streams: video + camera pose + inputs + world-state · one row / frame
ANNOTATION session_export: engine ground truth · 60 Hz · 4 synced POV
GROUND TRUTH
Depth Normals Albedo Optical flow
Video for review · 32-bit EXR for training

Ground truth you can't film

Footage has to be labelled by guesswork. We build the world instead — so every label comes out exact, by construction.

Because the environment is ours, every channel is read straight out of the renderer instead of estimated after the fact. The sample below is one scene, delivered five ways, frame-locked.

BEAUTY — THE SCENE AS RENDERED
Same frame, five ways 0.00 s

This sample was built to a customer's specification: they told us which channels they needed, we designed the environment and produced them. The channels, the environments, the camera behaviour and the volume are all things we set at the start — not things we hope to find in existing footage.

DEPTH The exact distance to every point in the frame.
NORMALS The exact orientation of every surface.
ALBEDO True material colour, with the lighting removed.
OPTICAL FLOW How every point moves, frame to frame.
Book a Demo →

CONVERSATIONAL REASONING

BETA

Pleiada MindSets

Human reasoning data from controlled conversational experiences.

Pleiada MindSets is our conversational reasoning data layer. We design controlled avatar-based scenarios where real humans interact with AI characters, revealing how they interpret context, reason through ambiguity, respond emotionally and make decisions.

SAMPLE
Conversational Reasoning Adaptive AI character Pleiada Labs
Audio · transcript · reaction signals · reasoning trace
48 kHz AUDIO
Word-level TIMED TRANSCRIPT
Frame-synced REACTION SIGNALS
Turn-aligned REASONING TRACE
REASONING TRACE LIVE turn-aligned · word-level
AIt 968.2s "Which of the two options would you prioritize if time were short?"
reasoning intent:"probe_tradeoff" · hypothesis:"user_uncertain" conf:0.71 · strategy:"ask_clarifying_question"
HUMANt 970.5s "If that's true, then the second option makes more sense."
SIGNAL prosody pause:1.2s · pitch:↓ → hesitation:0.68 @ t 971.8s
HUMANt 972.4s "Okay, but what happens if it fails halfway?"
SIGNAL reaction:"reconsideration" conf:0.68 @ t 973.3s
reasoning detect:"risk_concern" · strategy:"reassure + concretize fallback" · tone:"calm"
AIt 974.1s "Then we fall back to the manual path — you'd lose two minutes, not the session."
SIGNAL reaction:"confidence_rise" conf:0.52 @ t 975.9s
DECISION choice:"commit_option_2" · latency_ms:335 · confidence:0.81
AIt 977.0s "Good call. What tipped you toward the second option in the end?"
reasoning intent:"elicit_justification" · goal:"capture post-decision rationale" · strategy:"open_question"
HUMANt 979.6s "Honestly? Knowing there was a fallback. I stopped worrying about the worst case."
SIGNAL prosody rate:↑ · smile onset:0.41 → relief:0.74 @ t 980.1s
reasoning detect:"risk_aversion → resolved" · update:"user values safety nets over speed" conf:0.77
AIt 981.4s "That makes sense. Would you have decided faster if I'd mentioned the fallback first?"
HUMANt 984.0s "Definitely. Lead with the safety net next time."
SIGNAL reaction:"assertive_close" conf:0.83 @ t 984.8s
ANNOTATION session_summary:"fallback framing reduced decision latency" · tags:["risk_aversion","reassurance","order_effects"]
Book a Demo →

Canonical Format

One normalized format, ready for any pipeline.

Every session is delivered as a per-frame table (CSV or Parquet) synced to its video — one row per frame, identical column conventions across every title.

Camera-to-World matrix

Every camera frame carries the full 4×4 pose matrix in a single, consistent world convention — ready to convert downstream into whatever pose format your training pipeline expects.

⎡ Rx  Ux  Bx  Tx ⎤
⎢ Ry  Uy  By  Ty ⎥
⎢ Rz  Uz  Bz  Tz ⎥
⎣ 0   0   0   1  ⎦

Camera intrinsics

Pinhole intrinsics with every camera session — focal lengths and principal point in pixels, derived from each title's field of view and resolution. Extended distortion models supported.

    ⎡ fx  0   cx ⎤
K = ⎢ 0   fy  cy ⎥
    ⎣ 0   0   1  ⎦

Input data

Per-frame active keys, mouse buttons and high-rate mouse deltas — plus semantic in-game actions like "Sprint", "Jump" or "Reload". A per-session key-binding map ties raw keys to the player's actual settings.

input_keys    W|Shift
input_actions Sprint|Jump
mouse_dx/dy   -3.2 / +1.4
ON REQUEST

QA visualized.

On request, we run automated QA across a sample of any delivered batch and hand back a frame-synced Rerun (.rrd) file — video, camera pose and inputs replayed together in the viewer — alongside the scripts that generated them.

SAMPLED SESSIONS PASS AUTOMATED SELF-CHECKS
✓Orthonormal rotations ✓det(R) = +1 ✓Level horizon ✓Plausible camera height
Rerun
Get in Touch → Pleiada

The human behavior layer for AI.