Block Catalogue

Find a block by its job, then open its page for input/output shapes, parameter positions, and startup requirements. Availability depends on the platform and build options.

Choose by task

I want to… Start with
Try MOSAIC without hardware Clock → Sin Generator → Function
Prepare signals Filter, Crop, Resampler
Extract features from windows Sliding Window → Metric Extractor
Combine or select channels Joiner, Projector, Channel Selector
Capture examples for training Trigger and Trigger Buffer
Predict or classify Batch Predictor, Classifier
Exchange data UDP Client, Serial Sender, LSL

Read a block page

Type lists accepted configuration keys. Consumes and Publishes describe payload types and shapes; connecting two nodes does not automatically convert them. Parameters are positional and describe model-loading defaults. Some palette dialogs expose fewer fields or choose different defaults. Follow the configuration guide when editing JSON.

Windows, macOS and iOS run MOSAIC, but vendor device blocks can require a specific platform, SDK or Python environment. Read requirements before choosing hardware. A registered factory type may also be absent from the palette; its page identifies that distinction where applicable.

Analytics

Block Purpose
Metric Extractor Collapses a windowed matrix to one scalar amplitude/feature value per channel.
Online ICA Streaming Independent Component Analysis — incrementally learns an unmixing matrix that separates mixed channels (EMG crosstalk, motor-unit sources) into k independent components.
Online LDA Streaming Linear Discriminant Analysis — maintains per-class means and within/between-class scatter online, projecting into a k-dim subspace that maximises class separation.
Online PCA Streaming Principal Component Analysis — incrementally tracks the top k principal directions (Oja/Sanger update with periodic QR re-orthonormalisation) and projects each sample onto them.
Supervised ICA Online ICA subclass that additionally stashes each projection into labelled clusters for the scatter-plot card.
Supervised PCA Online PCA subclass that additionally stashes each projection into a colour-coded, labelled cluster set for the scatter-plot card.
Supervised UMAP Label-guided non-linear dimensionality reduction via umap-learn over pythonnet.

Devices

Block Purpose
BodyRig DLR BodyRig inertial motion capture.
Command Sender Fire-and-forget UDP command sender for ESP-based IMU nodes.
DLR ADC/BT Acquires analogue samples over a Bluetooth serial port.
Delsys Delsys Trigno RF wireless EMG source.
Hannes Hand Sink driving the Hannes prosthetic hand over a BLE serial dongle using the EMGEM protocol.
MCC DAQ Acquires analogue inputs through the Measurement Computing Universal Library.
Muovi Streams HD-sEMG / EEG (optionally plus IMU quaternions) from up to 16 OT Bioelettronica Muovi probes attached to a SyncStation, over an outbound TCP connection to 192.168.76.1:54320.
Muovi Single Probe The same decode as Muovi, but for one probe connecting directly over Wi-Fi with no SyncStation: this block is the TCP server, binding 0.0.0.0:54321 and waiting for the probe to dial in.
Myo Armband Tick-driven BLE source for the Thalmic Myo armband: dequeues one BLE EMG sample per tick and decodes 8 signed bytes through a 256-entry LUT.
Quattrocento Streams multichannel EMG/EEG/AUX from an OT Bioelettronica Quattrocento amplifier over TCP (default 169.254.1.10:23456), decoding the device frame down to only the inputs the user enabled.
SiFi Source for SiFi Labs BLE biosignal wearables (BioArmband / BioPoint), driven by an external sifibridge child process whose JSON stdout is parsed into EMG/IMU/ECG/EDA/PPG streams.
Stimulus Automated cue generator: walks a list of target activation vectors through a timed FSM (rest → rise → hold → capture → fall → next task) and publishes the current state plus interpolated target each tick, so downstream trainers know when to record.
WULPUS Bridges the Wulpus wearable ultra-low-power ultrasound dongle into the pipeline through its Python package (pythonnet), publishing A-mode acquisitions or assembled B-mode frames.

Flow Control

Block Purpose
Buffer Start/stop capture accumulator for batch training.
Channel Selector Per-channel enable mask driven from the UI.
Control Algorithm Maps a prediction vector onto prosthesis degrees-of-actuation via a runtime-swappable IControlAlgorithmStrategy.
Joiner Fan-in: concatenates the most recent value from every input into one flat vector, emitted once per arrival from a designated timer input.
Manual Control UI sliders as a degrees-of-actuation source.
Matrix → Vector Replays a buffered [timesteps × channels] frame one row per clock tick, turning windowed output back into a continuous sample-by-sample stream.
Mean Average Value Fixed single-feature extractor: MAV = (1/N)·Σ|x[i]| per channel over a window.
Projector Keeps only the listed channel indices, producing a reduced-dimension output.
Selector Unwraps a (label, payload) tuple and forwards only the payload, so downstream blocks expecting a bare Vector/Matrix can consume tagged output unmodified.
Sliding Window Accumulates incoming rows and publishes a BufferSize × Channels matrix every Stride rows, giving downstream feature blocks the overlapping windows they need.
Slope Sign Changes Counts slope sign reversals per channel per window — a standard EMG time-domain feature.
Stim Trigger Adapter converting a Stimulus block's FSM state tuples into the Buffer start/stop protocol, so a scripted stimulus schedule can drive capture without user clicks.
Stream Trigger Proxies a continuously-varying Vector source (dataglove, BodyRig, SinGenerator) into a Trigger Buffer as per-frame regression targets, republishing whatever arrived most recently.
Switch A/B selector — forwards whatever the currently active one of two upstreams publishes and drops the other.
Trigger Manual label source for supervised training.
Trigger Buffer Captures labelled training segments, pairing every data frame with whichever target the trigger published most recently — one storage shape for both classification and regression.
Waveform Length Fixed single-feature extractor: WL = Σ|x[n] − x[n−1]| per channel per window, capturing amplitude and frequency content together.
Zero Crossings Counts sign changes through zero per channel per window — a cheap proxy for dominant frequency content.

Machine Learning

Block Purpose
Batch Predictor Multi-output regression fit in one shot over every labelled segment held by an upstream Trigger Buffer, then predicting continuously.
Classifier Multi-class gesture classifier with two training modes: batch retraining from a Trigger Buffer, or per-sample incremental updates driven live by a Trigger.
Hybrid Predictor Trains on discrete gesture labels like a classifier but outputs a continuous control vector — the probability-weighted blend of the per-class target vectors defined by the Trigger.
Incremental Predictor Online multi-output regression: updates on every sample while a Trigger holds a target vector, and predicts continuously otherwise.
Python Predictor Classification delegated to a Python model over pythonnet — the block ships windows of data into a user-supplied Python Model class and republishes its probability output.
Python Predictor (Regression) Continuous multi-output regression delegated to a PyTorch model over pythonnet — windows of multichannel EMG in, a smoothed joint/DOF activation vector out.
Ultrasound Classifier ResNet-18 over pythonnet: one B-mode ultrasound frame in, a softmax gesture-probability vector out.

Signal Processing

Block Purpose
Adaptive Filter First-order IIR low-pass whose cutoff is recomputed every sample as fc = exp(Offset + Deriviate·\|dx_filt\| + Magnitude·\|x\|).
Crop Slices a sub-range out of a vector or matrix — trimming an A-mode line to a region of interest, discarding leading metadata samples, or keeping a subset of channels.
Cross Correlation Compares two channels to estimate their relative lag.
Depth Filter The same FIR/IIR designs as Filter, but applied along each row (the depth axis) with the state reset before every row, so frames are independent.
FFT Single-sided FFT spectrum per channel from a windowed matrix, scaled as magnitude, normalized amplitude, power, or dB, feeding a live SpectrogramMonitor.
Filter The general-purpose time-domain frequency filter: an independent IIR (Butterworth) or FIR instance per channel, state carried across samples, parameters adjustable at runtime.
Function Applies one element-wise mathematical map to every element of the stream — rectification, offset, scaling, power, clipping, or a soft threshold.
Negate Flips the sign of each channel in a vector.
Resampler Sample-and-hold rate converter.

Streaming

Block Purpose
Blender Arm Converts a control-algorithm actuation dictionary into a 9-element normalised command vector and ships it as an ASCII JSON array over UDP to the Blender-side Python script.
Clock Master timing source: a dedicated highest-priority background thread publishes a monotonic timestamp at DesiredRate Hz, driving every synthetic pipeline.
LSL A Lab Streaming Layer endpoint that is either an outlet (pushes pipeline data onto the network) or an inlet (pulls a resolved network stream in), selected by Params[0].
Mock Ultrasound Synthetic ultrasound frame source for exercising the classifier/trigger-buffer pipeline with no hardware, emitting deterministic class-conditional grayscale patterns plus Gaussian noise.
ROS A rosbridge_suite bridge over WebSocket (ws://IP:9090): subscribes to a topic and publishes each message into the pipeline, or advertises a topic and forwards pipeline vectors/matrices out as std_msgs data arrays.
Serial Sender Writes formatted numeric samples to a serial port with a background writer.
Sin Generator Synthetic multi-channel sine source: on each clock tick it evaluates A·sin(2πft + φ₀ + i·Δφ) per channel, using a jitter-free internal sample counter rather than the incoming timestamp.
UDP Client Bidirectional UDP endpoint: binds a local receive port and publishes every parsed datagram downstream, and — if an upstream and a remote endpoint are configured — also serialises upstream vectors out.
UDP Control Client Sends control vectors with proportional and derivative gains and receives feedback.
UDP Sender Sends pipeline data to Unity's Streamlined Input Manager over UDP using its binary framing — count, value/data/subtype bytes, 8-byte timestamp, payload, UInt16 counter — choosing the serialisation by the upstream block's runtime class name.

Tests

Block Purpose
TAC Test Target Achievement Control test (Simon et al. 2011): scores how well a user drives a predictor onto Stimulus-supplied target postures.

Add a block

Follow Build Your Own Block, then Catalogue Registration. Add a reference page alongside its model and palette entry.

Search documentation

Start typing to search the documentation.