animationsight 0.7.2

Animation review built exclusively for AI agents: a clip in, exact motion measurements + evidence renders out.

animationsight

Animation review built exclusively for AI agents. A clip in; exact motion measurements, located findings, and renders of the frames that are wrong out.

An animator judges a clip by watching it. An agent cannot watch anything — and the defects that matter are the ones that hide from watching anyway: a foot that slides 2 mm per frame, a single-frame pop in a 4-second take, a toe 40 mm through the floor for 5 frames. Those are all exact arithmetic. So do the arithmetic.

clip.bvh  ->  animationsight inspect  ->  report.json + flagged frames  ->  fix  ->  diff

What it measures

kinematics	per-joint velocity, acceleration, jerk (mm/s, /s², /s³) and true geodesic angular speed
contacts	plant/lift events per foot, with frames and times
foot sliding	drift while planted: total slip, peak speed, worst frame — the classic defect
balance	COM (real segment masses) vs the support base, per frame, with airborne frames separated
ballistics	effective gravity per airborne span from a parabola fit — floaty/heavy flights, quantified
penetration	any joint through the floor, with depth and frames
smoothness	discontinuity events, clustered: "pose snap" (many joints, one frame) vs "joint pop" (one bad key)
loops	seam gap judged against the clip's own per-frame motion, convention-aware; --kind oneshot declares the intent

Findings carry where (frame, joint, magnitude) and a try:. Renders mark the offending joint and draw the COM, the floor and the support. Everything is deterministic: same clip, byte-identical report.

Quickstart

pip install animationsight

animationsight inspect walk.bvh --unit cm --up y
# -> out/report.json
# -> out/frames/frame_0022.png   (the flagged frames, offender circled)
# -> out/track_foot_speed.png    (time series, slides marked)

animationsight diff before.bvh after.bvh    # prove the fix
animationsight track walk.bvh LeftFoot      # one joint, every frame

Or as a Claude Code plugin:

/plugin marketplace add VortexJer/AISight
/plugin install animationsight@aisight

(the plugin carries the skill; the CLI it drives still comes from pip)

Exit codes: 0 ok, 1 bad clip, 2 a FAIL-level finding.

Proof it works: known ground truth

examples/01-walk is synthetic on purpose. Real mocap has unknown defects; a clip whose defects are known by construction is the only kind that can prove a measurement is right.

make_clips.py builds a walk by authoring the foot's world trajectory (stance planted, swing arcing) and solving the leg by inverse kinematics, then writes two files: a clean reference, and the same walk with three injected defects of exact magnitude.

injected	measured
right foot dipped 40 mm, frames 22-26	[FAIL] 'RightFoot_end' goes 40.0 mm through the floor — worst at frame 22, 5 frames
left arm displaced 26° on frame 47 only	[WARN] 9 single-frame acceleration spike(s); worst on 'LeftForearm' — frame 47, robust z 102.5
root travels 15 % faster than the stride	[WARN] 'LeftFoot_end' slides while planted: 130.5 mm over 58 frame(s)

And the clean reference reports OK, exit 0 — no findings. Both directions are asserted in tests/.

real committed output: frame 22 with the penetrating foot circled below the floor line · the foot-speed track with the sliding frames marked

Before / after: one turn of the loop

examples/02-jump is the defect nobody can see and everybody can feel: a jump whose flight falls at 0.58x gravity. Every still frame looks fine; the parabola fit does not care.

flight: frames 27..45 (0.6333s, apex +266.2 mm) -> 0.58x gravity
[WARN] ... at 1 g that apex takes 0.47s of airtime
       try:   T = 2*sqrt(2h/g): shorten the airtime or raise the apex

Apply the try: line, re-inspect, and prove it:

animationsight diff jump_floaty.bvh jump_fixed.bvh --kind oneshot
  flight 0: 0.58x gravity (0.6333s) -> 1.064x gravity (0.4667s)
  GONE [floaty-flight] flight at frames [27, 45] falls at 0.58x gravity ...

the arc sheet inspect writes for every flight: ghosted poses, the measured COM arc (red, one dot per frame) vs the 1 g reference (green, dashed). Left: floaty — the red arc overshoots the reference by a third of the jump. Right: fixed — the arcs coincide.

Blind vs measured: the parkour vault

examples/03-parkour is the full study: a 240-frame run → vault → land → turn → settle sequence authored twice — once by a cold-context agent with no tools at all (numpy and arithmetic, one shot), once through this tool's loop. The blind clip is genuinely strong — hand-solved vault ballistics that measure 1.03x g, IK-pinned feet, zero sliding — and still ships what no eye can see: a stride flight at 0.47x g, a turn step whose "flight" is really IK overreach lifting the planted foot, a 116607 mm/s² knee pop at the landing. The after clip audits OK, zero findings, and the diff proves each fix did what it meant.

left: the blind one-shot — competent enough that its defects are exactly the invisible kind · right: the after clip — same choreography, physics that measures clean

the stride flight's arc sheet — blind: the measured COM arc (red) runs flat at 0.47x g while the 1 g reference (green) falls · after: 1.0x g, the arcs coincide

Four bugs this found in itself

Worth stating plainly, because they are the same class of bug the tool exists to catch:

1. The first walk generator did not plant its feet. It was built from sinusoids, which look like a walk and are not one — every frame slid. animationsight caught it immediately (min foot speed 81 mm/s, never zero), which is why the generator now uses IK and why test_clean_clip_plants_its_feet exists.
2. The floor was inferred from the clip's lowest point — so the injected penetration defined the floor and made itself invisible. Now it is the 10th percentile of the per-frame lowest point: a defect is rare by nature and stays below it. (test_a_penetration_cannot_define_the_floor)
3. The author's own jump was floaty. While dogfooding, the ballistics metric measured the freshly-authored demo jump at 0.68 g — written as an "arbitrary nice parabola" and shipped without a second look. It is now examples/02-jump, defect preserved, with the fix beside it.
4. diff printed Nonex gravity when a flight exists on only one side — which is precisely what a good fix produces (the parkour turn step stopped being a "flight" once its planted foot stayed down). It says no flight now, regression-tested.

Scope, honestly

Reads BVH. Measures kinematics, contacts, balance, smoothness, loops.

Does not do: character rendering, blend shapes / facial rigs, retargeting, physics simulation, or editing clips. It reviews; you fix the clip in the tool that made it.

For 3D geometry — parts, assemblies, robot URDF, collision sweeps — see solidsight, the sibling tool this shares its philosophy with.

The Claude Code skill

skills/animationsight/SKILL.md ships inside the pip package and installs itself into ~/.claude/skills/animationsight on the first run.

Leaving: animationsight uninstall removes this tool's skill and package. pip install aisight && aisight uninstall removes the whole family — every skill, every package and the plugin marketplace. See aisight.

License

MIT