Train and publish
In Teachable Machine, give each class a useful name such as jump, card, or clap. Export the model and copy its share URL.
Complete block reference (48 blocks)
TurboWarp extension · user guide
Turn movement into project input.
The extension connects a Teachable Machine Pose, Image, or Audio model to TurboWarp. Your camera frame or microphone audio becomes label probabilities that scripts can read as labels, scores, and Boolean conditions.
INPUT
↓
PROBABILITIES
↓
PROBABILITIES
Current label
jump0.91
Kamishibai DSL 4.0 bundle
Open the Teachable Machine manual from the combined palette
Click the documentation button immediately below the Teachable Machine member
heading. The bundled blocks retain the model URL, recognition, preview, label, score,
threshold, and lifecycle behavior described here. The TurboWarp Teachable Machine 3.5.1 package line keeps the public lifecycle
opcodes startRecognition, stopRecognition, and
isRecognizing, with no legacy aliases. Configurable SVG pose overlay blocks
remain available in pose mode, while image mode uses label scoring without keypoints. Latest-needed model initialization
and cancellation controls remain host APIs documented under Composition API.
Quick start
From model to motion in four steps
You need a published Teachable Machine Pose, Image, or Audio model, the matching camera or microphone permission, and network access. Load the downloaded tm.js from the custom extension dialog’s Files tab with Run without sandbox enabled, then place these blocks in order.
Set the model URL
Paste the model folder URL. A missing trailing slash is added automatically.
Start recognition
Allow camera or microphone access. The extension starts the required input and loads the model automatically if needed.
Use the result
Read the top label, test one class, or use a confidence value to control your project.
when green flag clicked
set model URL tohttps://teachablemachine.withgoogle.com/models/…/
start recognition
forever
if label is jump with confidence at least0.75
How it works
One input, several useful answers
While recognition is running, the extension repeatedly samples the selected camera or microphone input, runs the model in the browser, and stores a probability for every class in your model.
Camera framelive image
Pose estimatebody keypoints
Class scores0 to 1 each
Blockslabel · score · test
current label is the class with the highest probability. confidence is that class’s score, rounded to two decimal places. confidence of [NAME] reads one named class.
The preview sits on the stage
The preview is a camera canvas placed over the TurboWarp stage. It does not become a sprite and does not change model input when hidden. It is mirrored by default, and can switch between mirrored and unmirrored while the camera is running without changing recognition input.
Version 1.11.0 overlays the 17 recognized PoseNet joints and 12 standard bone connections as SVG by default. Blocks control visibility, each joint's color, opacity, and radius, shared bone color, opacity, and width, the minimum confidence, and independent confidence scaling for those visual properties.
The extension requests a normal Canvas2D context for its owned camera canvas. Physical-camera measurements of the 320x240, one-draw/one-read path did not show a repeatable end-to-end benefit from willReadFrequently; the hint helped a different read-heavy condition with four or more reads per draw, but could make video drawing more expensive. Chrome may therefore emit its Canvas2D readback warning during CPU inference. Other TurboWarp canvases are unchanged, and hosts do not need a global Canvas or Console patch.
camera
top-lefttop-rightbottom-leftbottom-rightcenterfull-stage
Start and stop are explicit
start cameraopens the camera and displays the preview.start recognitionstarts the required input and loads the model when necessary.stop recognitionclears live results, stops audio listening, and leaves the camera available.stop camerastops recognition, releases media tracks, and removes the preview.hide camera previewonly changes what is shown; recognition can continue.
Composition hosts
Prepare only the model the story still needs
The composition API lets an application host cancel obsolete model demand while camera startup continues independently. The first inference waits for both the active input and one registered model.
Independent work and synchronization
Descriptordecode + SHA
PoseNetverified bytes only
Registeredjoins classifier
SHA checks for separate PoseNet files run in parallel. Shared descriptor verification overlaps classifier loading, but unverified bytes never reach TensorFlow. Camera getUserMedia() and video.play() may run alongside the complete model path.
Latest-needed wins
Activecancel at safe boundary
Pendingnever starts
Neweststarts after cleanup
Create the composition with modelInitializationPolicy: 'latest-needed' and pass an AbortSignal to registerPoseModel(). At most one heavy request is active and one newest request is pending. A cancelled model is never published to the registry.
Choosing a result
React now—or smooth motion over time
Default builds expose immediate probabilities. Builds that enable the optional temporalPoseScoring feature also expose accumulated scores for steadier decisions.
Current confidence
Best for quick gestures. The answer follows every frame, so it reacts fast but may flicker near a decision boundary.
confidence per framethreshold
Use label is [NAME] with confidence at least [THRESHOLD]?. Thresholds are clamped to 0–1; 0.75 is a practical starting point.
Accumulated score
Best for sustained poses. Evidence is added over elapsed time while older evidence decays, reducing one-frame changes.
accumulated scorethreshold
previous × decayseconds + probability × accumulation × seconds
The reporter returns an empty label until the best positive score reaches the threshold. Hidden-tab time is excluded. Stopping recognition or resetting clears all accumulated scores.
Block guide
Pick the block for the job
Reporter values are empty or zero before the first usable recognition result and after recognition stops. Boolean status blocks are useful when coordinating startup.
Model setup
Configure and check the Teachable Machine model.
set model URL to [URL]changes the model and clears the previously loaded one.load modelfetchesmodel.jsonandmetadata.json.model is loaded?reports readiness.Teachable Machine versionreports the extension build.
Compute backend
Choose how much of the machine recognition may use.
set compute mode to [MODE]·auto(default),webgpu,webgl,wasm, orcpucompute modereports what the project asked for.active compute backendreports what the browser actually gave.
auto tries WebGPU, WebGL, WASM, and CPU in that order and keeps the first one that really computes. A named mode is a preference: when the browser cannot provide it the same order takes over, and last error explains the substitution. Change the mode before loading a model—switching releases a model loaded from a URL so it can be rebuilt on the new backend.
Camera and preview
Control capture separately from recognition.
start camera·stop camera·camera is running?refresh camera listdetects available video inputs.set camera to [CAMERA]· default, front, back, or a detected devicecamera count·camera device ID·camera device nameshow camera preview·hide camera previewcamera preview is visible?set camera preview opacity to [OPACITY]clamps opacity to 0–1.set camera preview position to [POSITION]set camera preview to [MIRRORING]·mirrored(default) orunmirroredcamera preview mirroringreports the current setting.set pose overlay [VISIBILITY]·pose overlay is visible?set [PART] joint color [COLOR] opacity [OPACITY] radius [RADIUS]set pose bone color [COLOR] opacity [OPACITY] width [WIDTH]set pose overlay minimum confidence to [CONFIDENCE]set pose [PROPERTY] confidence scaling [STATE]
Camera labels may appear only after permission is granted. Device IDs are browser- and machine-specific. Switching a running camera preserves the model, recognition state, and preview settings; a failed switch attempts to restore the previous camera.
When TurboWarp-Camera-Source is loaded, this extension leases the camera it shares under the pose role, so one device serves recognition and anything else reading frames. stop camera then releases the lease without stopping tracks other consumers are reading. Camera Source is optional: without it the extension opens and stops its own camera as before, and nothing in a project changes.
Recognition and results
Start the loop, read scores, and make decisions.
start recognition·stop recognition·recognition is running?current label·confidence·confidence of [NAME]label is [NAME]?uses a fixed 0.75 threshold.label is [NAME] with confidence at least [THRESHOLD]?
Timing and errors
Measure startup and show useful failure details.
camera startup time (ms)model load time (ms)first recognition time (ms)last errorkeeps the latest runtime error message.
Accumulated-pose blocks
Live confidence flickers frame to frame near a decision boundary; these blocks combine evidence over time so a brief misrecognition cannot flip a decision. They are in the palette by default (temporalPoseScoring): set accumulated pose accumulation [ACCUMULATION] decay [DECAY], set accumulated pose threshold [THRESHOLD], reset accumulated pose scores, accumulated pose, accumulated score, and accumulated score of [NAME]. Decay is the fraction retained after one second (0–1), and a decay change takes effect the next time recognition starts.
Troubleshooting
When the result is not what you expect
Check last error first. Most setup failures come from permissions, an incorrect model URL, blocked network requests, or loading the extension in the sandbox. A sandboxed load is silent in the editor: the dialog keeps waiting and only the browser console shows TM must run without the extension sandbox. Load the file from the Files tab, never from the URL tab.
start camera also needs the page to have been interacted with once. Browsers refuse video.play() without user activation, reported as NotAllowedError: play() failed because the user didn’t interact with the document first. Clicking the green flag counts, so an ordinary project is fine; a project that starts on its own must wait for a button press.
The camera does not start
Allow camera permission for the TurboWarp page and make sure another application is not exclusively using the camera. The page must run in a secure context (HTTPS or localhost). Then run start camera again.
The model does not load
Use the published Teachable Machine model folder URL—not a link to the editor or an individual JSON file. The extension appends model.json and metadata.json to that URL. Confirm that the network allows jsDelivr and Teachable Machine model hosting.
The label is unstable
Add more varied training examples, keep the whole body visible, improve lighting, and test a threshold such as 0.75. If your build exposes accumulated-pose blocks, use them for a pose that must be held for a short time.
The preview is missing but recognition works
Run show camera preview, set opacity above zero, and choose a supported position. Packaged projects with a modified stage DOM may not support preview attachment; last error explains stage-detection failures.