01
Observe
Street imagery, satellite tiles, video. Ordinary RGB, no dense active sensing required.
Open spatial perception framework
Perception that checks itself.
An open framework that turns ordinary imagery into confidence-aware spatial representations, and says so when it does not know.
Agents can read the internet. They cannot read the street.
Today's AI systems understand documents, interfaces, APIs and text. What they lack is a persistent, structured, uncertainty-aware representation of the physical world: what stands where, how large it is, what the evidence for that belief is, and how much to trust it.
Capturing the planet with dense active sensing does not scale. Trusting raw visual reconstruction does not either: depth models drift, detectors hallucinate, and maps go stale. The missing piece is not a bigger model. It is a verification layer between imagery and belief.
Spatial intelligence scales when reconstruction is continuously cross-checked.
The core innovation is not a single perception model. It is a self-correcting framework: form a spatial hypothesis, compare it against independent semantic, geometric, physical, temporal and geographic references, estimate uncertainty, surface contradictions, and decide when additional evidence is required. A single reference can be wrong. Multiple partially independent references collectively constrain the scene.
Passive reconstruction
- Accepts model output as scene truth
- Scale drifts with no external anchor
- Errors propagate silently downstream
- Uncertainty is hidden or discarded
- Fails quietly on unusual scenes
Self-correcting perception
- Treats model output as a hypothesis
- Anchors scale to known dimensions, standards and maps
- Surfaces contradictions, never averages them away
- Ships confidence and provenance with every figure
- Abstains and requests evidence when uncertainty is high
From pixels to defensible belief.
02
Reconstruct
Detection, segmentation, metric depth, optionally SLAM or splatting, through model-agnostic adapters.
03
Retrieve references
Object dimensions, human and vehicle proportions, road standards, maps, temporal observations, optional sensor calibration.
04
Check consistency
Agreement between independent references raises confidence. Contradiction is recorded with both readings intact.
05
Quantify and decide
Uncertainty is estimated per figure. Too high, and the system requests more evidence. This step loops back.
06
Emit
Task-relevant geometry with provenance, bands instead of false precision, explicit unknowns. Solyzer consumes it today.
How references argue, and what happens when they disagree.
Estimating the scale of one street scene from a single photo. Each reference implies a scale correction for the reconstructed geometry. No single reference is trusted alone.
ReferenceKnown valueImplied scaleStatus
Motorcycle wheelbase1.29 m0.98Consistent
Rider height, seated1.35 m1.02Consistent
Road width, class prior3.0 m per lane0.96Consistent
Storefront doorway0.95 m1.31Contradicts
Three independent references agree within 6 percent. The doorway disagrees by 31 percent, so it is flagged as probably nonstandard rather than allowed to corrupt the scale. The confidence band stays moderate, and if a second viewpoint exists, the system requests it before narrowing.
Architecture
Modular layers, no vendor lock.
Every layer is replaceable. Detectors, depth models, reconstruction methods, databases and simulators plug in through defined interfaces, so the framework strengthens as the ecosystem around it improves.
L1
Perception adapters. Detection, segmentation, tracking, pose, metric depth, SLAM or splatting. Any model that emits hypotheses can drive the framework.
L2
Spatial memory and scene graph. Objects, surfaces, relations and observations, persistent across time, keyed to the ground.
L3
Reference retrieval. Object dimension tables, GIS and maps, product and CAD data, road and building standards, physical priors.
L4
Contradiction and uncertainty engine. Consistency checks across references, per-figure confidence, explicit abstention, degenerate-output detection.
L5
Simulation and sensor calibration. Optional. Synthetic priors for controlled testing; sparse LiDAR or depth anchors where available, as calibration rather than dependency.
L6
Confidence-aware world representation. The output contract: task-relevant geometry, provenance in-band, bands not false precision, unknown stated as unknown.
Early numbers from the reference implementation.
Measured on one consumer laptop GPU, running live inside Solyzer's perception stack. Reported with their limits, because that is the point.
6.55m
Road width from photos, 376 samples. An independent structural model said about 7 m. Two methods, blind to each other, 7 percent apart.
12,000+
Segmented object shapes across four districts, each with polygon contours and per-object confidence.
469vs 0
Trees observed by vision in a district where the map recorded none. Observation and mapping are separate layers, compared, never blended.
lowtrust
The self-calibration verdict of the first measurement run, published as-is. Vehicle sizes did not yet validate the depth scale, so the artifact says so.
Known open problems: per-scene camera height without mount metadata, canopy clusters versus individual trees, calibration transfer across districts, degenerate detector output on dense fabric.
Physical AI
Every machine maps the world alone. The map dies with it.
The cost of a wrong belief changes when AI gets a body. A hallucinated fact wastes a query and a retry. A hallucinated meter puts a gripper through drywall, a wheel off a curb, an inspection report in court. Physical AI cannot run on plausible. It needs world state that can carry the weight of a machine, with uncertainty priced in.
Today the industry answers that risk with hardware. Each machine carries its own dense sensor stack, rediscovers the same street on every pass, and takes what it learned to the scrapyard. The cost is paid per machine, per trip, forever. A shared, verified prior inverts the economics: the world is sensed once, checked against many references, and every machine after the first starts from a map instead of from zero.
L2
A world it can inherit. Spatial memory keyed to the ground means a machine would arrive with a map that was already argued over, so its own sensors go to what changed, not to resurveying what held still.
05
Knowing when to look again. Confidence per figure is a control signal, not a caption: trust the prior where the band is tight, slow down and gather more evidence where it is not.
L5
Sensors that pay forward. When a machine does carry LiDAR, L5 is built to take its returns as calibration: one pass would tighten the bands for every consumer after it, so perception compounds instead of resetting per machine.
06
The same contract, faster clock. Step 06 already has a consumer: Solyzer's digital agents act on its output contract now. An embodied system would read the identical output, only with far less time to decide.
To be plain: no robot, drone or vehicle runs on this today. Embodiment is a research direction, and the roadmap below labels it as one. The bet is specific. Trustworthy world state is the bottleneck, and it is cheaper to solve once, shared, than per machine. Solve it, and navigation, inspection, risk analysis and machines that pool what they learn all sit downstream.
Roadmap
Achievable stages, honestly labeled.
Exists
Camera-based reconstruction improved by semantic and geometric cross-checks: measured road geometry, segmented objects, in-scene self-calibration, demand-driven coverage
Building
Persistent uncertainty-aware spatial graphs; orientation-invariant calibration anchors; field-measured ground truth and public benchmarks
Building
Deep integration with Solyzer's explorable world model, where digital agents already inspect places, plan routes and evaluate environments
Research
Digital agents navigating and reasoning over the world graph virtually, with active evidence requests
Research
Fusing world priors with live RGB, LiDAR, radar, IMU and depth on drones, vehicles and robots: the prior gives context, live sensors correct the present
No claims of cameras becoming LiDAR, complete Earth understanding, or solved embodiment. The defensible claim is narrower and stronger: cross-reference verification reduces dependence on dense active sensing, and confidence-aware representations are what agents can actually act on.
Open enough to become a standard.
Opens here: the framework and its schemas, the scene graph format, plugin interfaces, reference pipelines, sample implementations, evaluation tools, benchmarks, and the failure cases that shaped them.
Lives in Solyzer: the continuously updated world graph, curated reference databases, production calibration, coverage, and the explorable product that agents and people use today.
What developers can do today: read the architecture, follow the reference implementation as it lands in this repository, and consume the live deployment through Solyzer's open API and MCP tools, where agents already query measured geometry, observed objects and coverage on demand.
Solyzer is powered by Hudhudbird Perceptrons.
Hudhudbird is the perception and verification layer. Solyzer is the world: an explorable, continuously updated digital twin where people, AI agents and future physical systems read the same ground. The framework opens here so the field can build on it. The world model ships there so anyone can use it today.
The name comes from the old story of the hudhud, the scout of Sulaiman's court: the bird that flies to a country the king cannot see and returns with a report he can act on. Go look at the world. Come back with reports you can defend.