GaitAI Insights

Ideas in motion.

Research, engineering and perspective on how machines understand human movement. Every story can be explored, not just read.

12 stories · Latest Oct 14, 2026

Drag to add observations1 observation
A population distribution beside one person's repeated observations forming their own narrow band, with the same latest reading marked against each.

What Is a Personal Movement Baseline?

A population reference says where a person sits among strangers; a personal baseline says whether they have changed. What a baseline is made of, how many observations it needs, when it moves, and what it can and cannot say.

Mobility8 min read

Read the story
Drag to break a linkAll sound
A chain of components from camera to operator, with one link broken while the model's link stays sound.

A Good Model Can Still Be a Bad System

A model can be right on every frame while the system around it fails: dropped frames, stalled inference, a network that swallows an alert, an interface that hides it, a person who has stopped trusting it. Model quality is one link in the chain.

Engineering8 min read

Read the story
Drag to strip the frameRGB
A figure stripped from a full image to a silhouette to a skeleton to joint trajectories, with the identity cues that remain at each stage.

Can a Skeleton Still Reveal Identity?

RGB, face removed, silhouette, skeleton, joint trajectories: each representation removes some identifying cues and keeps others. Where identity disappears depends on storage, combination, persistence and purpose — not on the picture alone.

Responsible AI8 min read

Read the story
Drag to move the cameraSide (left)
A camera orbiting a walker seen from above, with the movement signals each angle keeps, weakens or loses.

Camera Angle Changes What AI Sees

A camera records a projection of a walk, not the walk. The angle decides which signals survive the flattening — knee flexion, stride width, timing, foot trajectory, body path — and a movement system has to be honest about which it has.

Engineering9 min read

Read the story
Drag to shift one sideLeft longer
Two gait cycles, left and right, drawn as bars of stance and swing, with the right side drifting out of alignment.

What Does Gait Symmetry Actually Mean?

Symmetry compares the left side of a walk with the right — but stance, swing and step timing can each be asymmetric on their own, nobody is perfectly symmetric, and what a change means depends on the baseline it is read against.

Movement intelligence8 min read

Read the story
Drag to reveal the frameAI view
A walking skeleton beside the camera frame it was read from, with one estimated joint sitting off the body.

When Pose Estimation Lies

Pose estimators almost always return a plausible skeleton. The failures that matter are the ones that still look like a person: a misplaced ankle, crossed legs, a foot that vanishes — and every gait measure downstream inherits them.

Movement intelligence10 min read

Read the story
Drag to decode the walkVideo
A filmstrip of captured frames resolving into a single movement signal.

From Walking Video to Movement Intelligence

A walking video looks simple. Turning it into reliable movement intelligence is not. Inside the pipeline from capture and pose estimation to gait features, sensor fusion and actionable outputs.

Movement intelligence8 min read

Read the story
Touch a readingOne movement
One gait signal branching into five readings: identity, mobility, recovery, risk and safety.

Your Walk Is More Than a Biometric

Gait has often been studied as a biometric signature. But human movement can carry information about mobility, recovery, functional change, risk and safety context as well.

Movement intelligence4 min read

Read the story
Drag to remove identityRGB
Four stages of abstraction — pixels, silhouette, pose, trajectory — with identity information falling away and movement information retained.

Can AI Understand Movement Without Identifying the Person?

Many movement-intelligence tasks do not inherently require identity. A privacy-aware architecture starts by asking what information the task actually needs.

Responsible AI5 min read

Read the story

Foundations

Five interactive essays on how human movement becomes machine-readable intelligence.

Choose a story to preview it. Read them in order, from a walking video to an audited multimodal claim — or start anywhere.

Begin the path
Tap a stream
Four input streams converging on one fusion node — one missing, one corrupted — with an uncertainty band leaving it.

When Fusion Looks Better Than It Is

More modalities and more complex fusion do not automatically mean better evidence. Five questions that expose whether a multimodal result is genuinely convincing.

Research5 min read

Read the story

Previewing Foundations 05: When Fusion Looks Better Than It Is. When does another sensor improve the AI — and when does it only improve the benchmark?

Explore by subject

Browse GaitAI insights

Explore research, engineering, mobility and responsible AI by topic.

Publication archive

Explore the archive

2026JulAugSepOct

Browse every article, research note, product update and engineering story by year and month.

View the archive