Machine learning · 2026
Ichne
Learning what wrist motion reveals about a shift.
Ichne explores a simple question: how much of a working shift can be understood from a wristband, without asking the wearer to stop and record every task? The current build focuses on the foundations—sensor quality, trustworthy labels and activity recognition—before extending into route reconstruction and reporting.
- Status
- Research prototype
- Domain
- Wearable research · Activity recognition
- Role
- Research architecture, data quality and machine learning
- Year
- 2026
A person moves. A sensor listens.
Movement through a room creates a stream of observations. The animation starts with a simple walk and a wearable sensor.
An illustrated research direction: from activity signals to spatial context. The floor map is a concept, not a measured reconstruction.
Stack
- Python
- scikit-learn
- PyTorch
- NumPy
- Supabase
- YAML
The activity-recognition system behind the CHARM band: a research prototype that turns wrist-sensor recordings into quality-checked data, activity predictions and a foundation for shift analysis.
A band, a collection system and a model
CHARM is the wearable; AutoChore provides the early collection and inspection tools; Ichne is the modelling system.
The work has moved from collecting labelled motion to a separate Python build with shared data contracts, quality checks and reproducible evaluation.
First, check what the sensor really recorded
Different devices report acceleration in different units and treat gravity differently.
Ichne records those differences explicitly and checks timing, gaps and sensor behaviour before a recording is used. Device acceptance tests bring the same discipline to each band before a study.
Measure learning without confusing it with the device
The feasibility study compares activity classifiers with simple baselines and holds out complete sessions or users.
The pilot is small and partly confounds wearer, device and activity, so cross-device generalisation remains unproven. That finding shapes the next collection study.
Build the tools needed for better data
The implementation now includes a fleet registry, label alignment tools, a browser labeler, prediction timelines and a demo dashboard using classifier output.
Read-only live quality checks and a plan-versus-actual joiner extend the path from recordings to useful shift review.
The next stage
Self-supervised pretraining, dock-anchored route reconstruction, floor maps and grounded narrative reports form the broader research programme.
The current prototype establishes the data and activity-recognition work those later stages depend on.
Pipeline
Current activity-recognition prototype
Stage Can reject
-
01
Record
Wrist-sensor data
-
02
Standardise
Consistent sensor units
Device differences recorded
-
03
Check quality
Usable recordings
Timing, gaps and device behaviour
-
04
Align labels
Study dataset
-
05
Classify
Activity predictions
-
06
Evaluate
Grouped test results
Session and wearer limits kept visible
-
07
Inspect
Prediction timeline
Behind the work
The skills
that came together.
Research
- Grouped model evaluation
- Device and wearer confound analysis
- Baseline selection and uncertainty reporting
Data
- Sensor canonicalisation
- Device acceptance gates
- Recording quality checks
- Label alignment and study tooling
Implementation
- Activity prediction and timelines
- Self-supervised pretraining scaffold
- Plan-versus-actual data joining
Project highlights
-
Quality before prediction
Timing, units and device behaviour are checked first
-
Activity-recognition prototype
Classifiers, timelines and a working demo dashboard
-
Study tooling
Fleet records, label alignment and live data checks
-
A clear research path
Routes, maps and narrative reporting are future stages