01 · e.g. 1X NEO, Figure
Humanoid and home robots
About to share hallways with older adults. They need to know a towel bar is not a grab bar.
A pre-clinical test ground for home-health AIopen prototype →The flight simulator
The flight simulator
for home-health AI
How do we test a home-health AI before anyone relies on it in a patient's home? HealthDojo combines synthetic safety scenes, walkable worlds, and traces showing what vision models look at, flag, or miss.*
Cluttered staircase · objects on the treads
synthetic · selected world
A selected world, expanded from a synthetic image with World Labs Marble into a walkable Gaussian-splat world. Recorded model walks use panoramic views from those worlds; this viewer does not score models.
Why now
AI is moving into the home
Three kinds of products are about to judge safety in real homes from what a camera sees. Each one should be tested before it meets a patient.
02 · e.g. Vision Pro, Meta Ray-Ban Display
VR headsets and smart glasses
Seeing the home through the user's eyes, in every room they walk through.
03 · discharge, monitoring, caregiving
Home-health and care apps
A wave of apps assessing homes from a phone camera, judging safety from pixels.
HealthDojo is a place to start: guideline-grounded synthetic homes where these models can be pressure-tested first, before any real-home study.
The loop · one hazard, followed through
From a guideline line to a model test
Follow one hazard, objects left on the stairs, from published guidance to a model walking the room. The rubric awaits OT/PT review, and the checks are automated.
- 01 · GuidelineCDC STEADI · Check for Safety
Stairs and steps: keep objects off the stairs; …
summary of source guidancePublished home-safety guidance, summarized with its source.
- 02 · Draft rubric
STAIR-03review pendingObjects on stairs- low
- Item against wall on wide stair
- medium
- Item on a tread
- high
- Items on multiple treads or at top step
Draft a hazard rubric from published home-safety guidance.
- 03 · Labeled image
STAIR-03Seed a hazard in a synthetic room and check its visibility.
- 04 · 3D splat world
Expand selected images with World Labs Marble into walkable Gaussian-splat worlds.
- 05 · Model test
Compare the seeded hazard with model flags and viewpoints. Recorded model walks use panoramic views from those worlds. Here, 4 of 6 recorded models flagged it within 8 steps.
In the photo benchmark, 23 of 23 models found this particular hazard. Harder ones follow below.
Drop a model in
Watch a model inspect a synthetic room
Saved replays show where it turned, what it flagged, and what it missed. They illustrate specific runs, not performance across the full photo benchmark.
MissedGPT-5.6 Sol · Bathroom · BATH-07
Bathroom miss: the model looks toward a loose bath mat but flags support fixtures instead.
More recorded runs and 3D worlds →Early results
One early finding
In one synthetic staircase image, 16 of 23 ranked vision models flagged a handrail that stops short. That is a single-scene finding, not an estimate of real-home accuracy. The draft rubric still awaits occupational and physical therapist review.
SEEDED · STAIR-02
Before the editIn this synthetic stair scene, the handrail stops short. Open the image and compare the saved model responses. This is one benchmark observation, not a prediction about safety in real homes.
Example scene · one synthetic image
16 of 23 ranked models flagged the short handrail
● Flagged it · 16
- Gemini 3.1 Pro
- Claude Opus 5.5
- GPT-6 Sol
- GPT-5.6 Sol
- Claude Fable 5.1
- GPT-6 Astra
- Gemini 3.8 Flash
- Grok 4.7
- Kimi K3
- GLM-5V Turbo
- GPT-5.6 Terra
- Claude Opus 5
- Claude Sonnet 5
- Llama 4 Maverick
- Qwen3-VL
- Gemma 3 27B
● Missed it · 7
- Grok 4.6
- Nova Pro
- Nova 2 Lite
- Mistral Medium 3.5
- Claude Haiku 4.5
- Nemotron Nano VL
- Mistral Large 3
At a glance
A small, inspectable prototype
23
vision models
ranked in the falls benchmark (at least 90% scene coverage); 3 more with partial coverage, not ranked
43
scored scenes
in that benchmark, including edited and clean rooms: 31 edits that passed the target-visibility check + 12 clean bases, from 47 candidates
35
hazard types
rows in the draft falls rubric, clinician review pending
11
selected walkable worlds
built from synthetic scenes
Open code and data
Inspect the evidence yourself
HealthDojo is an open prototype for testing how vision models judge home hazards. The first themes are falls after surgery and dementia home safety.