Skip to content

Motion, continued: every open motion issue, in full - #939

Open
nicolas-maman wants to merge 283 commits into
mainfrom
feat/motion-next
Open

nicolas-maman wants to merge 283 commits into
mainfrom
feat/motion-next

Conversation

@nicolas-maman

@nicolas-maman nicolas-maman commented Nov 4, 2026 •

Copy link
Copy Markdown
Collaborator

Every motion issue still open, implemented in full, in this one pull request, continuing #921 (merged at 61964ad): #853 and its milestones, #903, #907, #909 to #911, #922, #929 to #938, #940 to #974. 32 of them are met here and closed by hand so far (listed at the end).

On this branch so far (to 2026-11-04)

Since 2026-11-05 (pushed, CI green unless said)

Done locally, landing next (each after the full list)

In progress (one agent each, in its own worktree)

Order of work from here

The full plan is in #853, section "Road to the merge and to Siltreach".

  1. Land Crowd instances stepped on the GPU: a million figures with the CPU's share near zero #991 once its shaders are regenerated; character: walk and run as separate gaits with a timed change between them, not mixed by weight across a speed band #974 with metrics: a foot thrown faster than its clip has it reads as nothing (FOOT_THROW) #972, then Motion warping: a take's root bent onto a target point and facing #987 and mover: a runner asked to a goal inside its turning circle orbits it and never arrives (one ran off the plaza) #973 on it; Animation over the network: the graph's inputs and events, not poses #911.
  2. The plaza (M6 migration onto the animation pipeline; old writers retired #860, Plaza: walkers by the ramp and steps cross their feet 263 mm and pop joints at 112 to 165 rad/s #977, Plaza on characters: frame time 6.9 ms to 9.7 to 15.7 ms mean with 24 mannequins #980), the car fixes (Plaza: when the car ends up above a dummy, both fly off into space #982, Avoidance does not see a person lying struck: walkers walk through the body #979), the steering speed-up redone on the new mover.
  3. Phase C: Traversal: jump, fall and land, vault, mantle, ledges and ladders #986 traversal (on Motion warping: a take's root bent onto a target point and facing #987), M3 IK and constraints: two-bone with limits, FABRIK, full-body IK, foot placement, aim, hand IK (L3) #857's one foot system, Two-person scenes: handing over, carrying, grappling, talking #909, Get-up: an arm the protective reach leaves overhead is swept through the floor at the hand-over (seed 26: a finger 25.8 cm under); blocks the reach as default #1001, Faces: blend shapes, blinks, saccades, eyes that look, visemes #910's consonants. Phase B: Shared meshes: world bounds written into the mesh would race once bounds run on the job pool #999.
  4. M6 migration onto the animation pipeline; old writers retired #860's retirement of the old writers, then the merge gate.

Every motion issue, in full

#854 M0, ground truth

  • tests/fixtures/people, as the issue lists
  • the metrics module, as the issue lists
  • a viewer tool: the figure with a skeleton overlay, the clip list, scrubbing, two figures side by side, the metrics on screen
  • the old motion modules frozen: bug fixes only

#855 M1, skeleton, pose, clips, retargeting, root motion

  • skeleton: up to 256 bones, skinning's 96-bone cap lifted, a humanoid map onto the UE5 names
  • pose: flat local transforms; local and model space; blend, additive, mirror
  • clip: skeleton-indexed tracks with a sampling cursor; a root track; events; contact and sync markers found at import
  • retarget: chain retargeting with rest alignment, chain scaling, pelvis height, feet and hands kept on their contacts (UAL1 Rigify and 100STYLE onto the UE5 mannequin)
  • root motion moving the capsule, or held for in-place use

#857 M3, IK and constraints

  • two-bone IK with limits, soft reach and twist distribution in rig; motion's private copy gone (3dccf75)
  • FABRIK chains (spine; chain_solve, a cone per joint; no tail rig to test yet)
  • full-body IK with effector weights (body_new/body_set)
  • foot placement from contact markers (pelvis offset, lock, toe and heel, slopes and stairs): one foot system replacing the four
  • look and aim over several bones (look_solve, aim_chain)
  • hand IK on props: a bowl in both hands (Hand-backs between motion and the animator pop joints, sink a sole and slide the feet #933), a rifle with two hands

#858 M4, physics asset and physical animation

#859 M5, ragdoll and get-up

  • full ragdoll on a knockdown or death, keeping the animation's velocities
  • rest detection (motion.down; test_knockdowns, 100 of 100 up within 0.22 s of rest)
  • get-up from a pose snapshot into a matched clip, the capsule under the body, the clip warped onto it, inertialized back into the graph
  • get-ups clear of other bodies and of the ground
  • partial ragdoll for hits, blending back

#860 M6, migration

  • the plaza, the motion showcase and the editor's figures on the new pipeline
  • Siltreach told how to move, with the SHAs
  • eating with a bowl on the pipeline: the bites to the mouth over a held carry, as test_eat holds motion.eat
  • retired: per-node clips and string lookup, every direct writer of skeleton bones, locomotion's second foot lock, motion's keyed get-up, sit, kneel and lie, hold and eat

#861 M7, behaviours on physical animation

  • balance, catch steps, staggers, protective falls, crawl and writhe on L4, the animation as the reference, measured against L5 and shipped only where they look better

#862 M8, crowds and animation LOD

  • crowd pose banks baked from the graph's clip data (0cbba27)
  • update-rate LOD for distant figures (0cbba27)
  • hand-over between a crowd instance and the full graph by pose snapshot (0cbba27)
  • a cost per figure per frame at each LOD, against a 2,000-figure budget (1a268f4; 3.1 ms against 2 ms planned, still open)

#907 the generic pieces fighting is built on (rescoped: the moves, their order, timing and damage are the game's)

#903 the character component

  • the showcase's figures each set up in a few lines through it (all eight, 6d524e6); the README's animation example the real code (examples/people)
  • every check in test_anim_pipeline and test_ik_constraints holding through it (so far through it: walking, stopping, strafing, turning on the spot, looking, aiming, stairs, a ramp, seats; test_humanoid 15, test_use 4)
  • states, takes and transitions declared as data: ae3d.animgraph (test_animgraph)

#909 the generic pieces two figures need (rescoped: the scenes themselves are the game's)

#910 faces

  • morph targets imported and drawn on both renderers, their weights as pose channels (487dfff)
  • an ARKit-like face set mapped as the body's humanoid map is (e444397 maps it; no CC0 face with ARKit's 52 shapes exists to commit; VRoid's own set is mapped for blinks and vowels)
  • procedural blinks and saccades, eyes turned by the look IK (834e998, 97c2dc2)
  • visemes from a voice track (dddae7f: five vowels; consonants and recorded speech left)

#911 animation over the network

  • the graph bit for bit the same on two peers from the same inputs, across platforms
  • a remote figure under 150 ms latency with jitter and loss within 2 cm and 2 degrees at its feet and hands
  • bytes per figure per second measured

#922 the get-up after a trip at a run

  • reproduced in an ae3d test, and the figure stands

Part of

Closed by hand once met on this branch (the PR says Part of, never Closes):

nicolas-maman added a commit that referenced this pull request Nov 4, 2026
#921 merged without its changes note; this one says what it brought, with
the numbers it was measured at, and what #939 has added since.
… edge meets the world (#946)

A sweep's hit carried only the object, the ragdoll and the point, and an
edge passed through walls. A hit now names what it met (items.HIT_BODY,
HIT_STATIC, HIT_OBJECT), the ragdoll bone (physics.DynamicRayHit.bone,
ragdoll_bone_of_body), the edge's velocity there and how far past the
surface it went that step: what a damage model reads. The static world is
swept too (physics.ray_static_hit), once a swing.

test_items: the sword on a physical person names its pelvis, the edge at
50 m/s, 230 mm past the surface; a wall where the edge passes is met
once. Item against item (a parry) follows.
…wing (#946)

A weapon passed through another raised in its way. Every live item is
kept (g_items), and a swinging edge within ITEM_MEET (3 cm) of another
item's edge, held or free, is a hit of kind HIT_ITEM naming it
(hit_item), once a swing. What the swing does then is the game's:
items.swing_stop ends a whole-state swing into what played before it, a
swing over the gait into the gait.

test_items: a sword stood where the swing's edge passes is met as an
item, and swing_stop ends the swing there.
macos-2 ran out the 54 minutes of its Run CI on #939, cold (fc0c124) and
warm (62b7315), the leak check unfinished, with the suites the PR adds:
the leak tier is split over macos, macos-2 and macos-3. AETHER_REF
v0.882.0 (0.881's core benchmarks, 0.882's release of a string a failed
list add was handed).
…te's rate (#948)

A graph file's capture read a whole BVH, with no events, named by its
path, and a state had no rate. A capture line takes `from` and `to`
seconds, `event NAME SECONDS` (any number), `asf SKELETON` (CMU's ASF and
AMC, `fps` 120 by default) beside BVH; its take is named as the graph
names it, so an item's swing can name it. A state line takes `rate R`.

test_animgraph: CMU subject 02's walk read from ASF/AMC and cropped to
0.5 to 2.0 s is the take built by calls within 3e-8 rad, its two events
firing at their seconds into the crop over its rate of 1.5.
A hit carried the velocity of the sample whose ray met it, or of the
edge's middle for its own line and a parry: off by the blade's length
times its turn. A rigid edge's point moves as the lerp of its ends, so the
hit reads its ends' velocities at its own point along the edge.

test_items: the hit on the mannequin reads 51.87 m/s, the edge's own at
that point over the step the sweep saw (0% off; 6.9% before).
items.swing played a swing's take as a whole state with root motion off,
so Sword_Regular_A's 0.82 m step-in slid the feet under a figure that
stood, and the mover kept driving it. items.swing_ready makes the swing's
state and arms its sweep without playing it; character.swing plays it as
an act: the capsule moved by the take's root motion, the mover left alone,
the stand taken up where it got to. items.swing_at is swing_ready and a
play, back to what played before as it was.

test_items 19: moved by the take's root motion (0.8245 m) within 6e-8 cm,
the hand 2e-12 mm off its grip. The take glides (sampled with its root
motion its front foot slides 1.2 m, its back foot 21 cm): a toe skates
117 mm against the take's own 595 the first time from a stand, 11 mm
swung again.
A foot's contact weight was read off the new state alone, so a foot the
stand held down was let go on a transition's first frame and the fade
carried it across the floor: the first whole-body swing from a walk's
stand skated a toe 117 mm, the same swing again 11. Contact and plant
weights now ease through a transition as the pose does: from what was
drawn going in (an inertialized play) or from the old state (a
cross-fade).

A foot let go more than 3 cm from where its clip has it, while the clip
keeps that sole on the floor, shuffles there: its sole lifted 2.5 cm over
the floor, carried across while up, set down straight where it got to.
Eased along the floor it slid; lifted as high as a stride, a heel walking
out from under a bench went 22 mm into it.

test_items 20: the first swing from the walk's stand 11.06 mm, the same
swing again 11.06 (the take's own 595). Seats skate 0.52 mm (1.0 before),
a get-up after an act 2.9, the bench 4.1 mm; the motion suite's 25
programs pass.
Every block was on the default material (white, full specular, roughness
0.5): the ground shone like a wet road in the sun. The ground is a dark
asphalt now and the deck, ramp, steps and kerb a grey concrete, rough
(0.92), with no reflection and a low specular. On a still the open ground
reads one even tone (luminance 83 from its median to its 99th
percentile).
#946 cast a static ray beside each moving one, doubling a swing's sweep to
125 microseconds unoptimised, and CI's test_items ran past its 200 a
swing. physics.box_touches asks the world's bounding-volume tree once a
step whether anything but the holder (the static world, or another body)
touches the box the edge swept; the edge's rays are cast only then.

test_items 20: 70 microseconds a swing unoptimised (125), its hits as
before (the person's bone at 51.87 m/s, the wall, the raised blade).
…by the meter (#929)

The stop on the hand-built graph (halt by distance, its mirror by
transition_or, no matched entry: Siltreach's pacers' shape) was measured
only where its take marks a foot planted, so a foot the stop marks up and
drags along the floor went unread. It is now braked to a point from 0.86
and 1.86 m/s, let go at four moments a quarter second apart so the halt
comes on each foot, on a physics ground with the feet stepping, and the
meter reads every toe on the floor through the halt: 0.12 mm at worst
from the walk, 0.006 from the run. Each run goes on until it has stood a
second; braked from a run, a stop took longer than the 3 s it had.
100STYLE's stop shuffles a foot along the floor in its last settle step,
and the take marks it up for one frame. The lock let go on that frame and
caught it again on the clip's point the next, 36 mm away: the foot jumped
in one frame, from a walk and from a run, every stop. A lock caught again
one frame after it let go, its clip keeping that sole within TOUCH of its
floor, is caught on the point and pivot it held. Short steps that rise
2.5 cm for two or three frames (a seat's) are lifts, and caught anew.

test_anim_pipeline 99: through the stop and its settle steps, from 0.86
and 1.86 m/s on either foot, no toe on the floor moves more than 3.4 mm in
a frame (36). Filling the gaps in the takes' contact marks instead fixed it
too, but changed what motion matching reads and its changes turned joints
at 45.6 rad/s (14.1). The motion suite's 25 programs pass.
…raph

animator.transition_offset(a) is the largest joint offset, radians, the
last transition eased out of, and transition_seconds(a) its blend: a
check allows the easing's 1.875 x offset / seconds over the takes' own
fastest (#906, #937).

test_anim_pipeline 100: walked and run off from standing on the
hand-built graph, through the start clocked by distance and its switch
into the gaits, no joint faster than the takes' own plus the easing's
(72% and 83% of it). A guard: #937's race shows on the character's
walk-off from a seat, where its fix is still to come (the attempts and
their numbers are on the issue).

test_items: the whole-body swing waits until the figure stands, as its
check says; a second's wait left it braking in its stop, and a swing
begun mid-stop slides the stop's stepping foot (#967).
…ool (#935)

The UE5 mannequin on character.match_add over 100STYLE's neutral takes,
turned on the spot by 90 and then 180 degrees: its fastest joint 18.1
rad/s, under the takes' own 22.1 (plus the easing's 17.0), no sole into
the ground, a toe skating 0.0003 mm. Held.

Then walked to a stool 2 m ahead (stool_ual1.usable: the stool's slot, its
sitting takes read from UAL1's file whatever figure sits), faced to it,
sat on, stood up from and walked on. It sits, no sole into the ground
(held); measured and not yet held, so #935 stays open: the sit's enter
eases out of the matcher's pose 85 degrees off the take's first frame
(a toe skating 6.7 mm), and a matcher's change on the walk on, a 0.47 rad
offset over 0.2 s, turns a thigh 34.7 rad/s (the takes' own 22.1).
…rough a state's transition (#935)

The feet's contact weights eased through a transition from a value taken
in play alone. A matcher's change inertializes without play, so it eased
from a stale weight: walking on from a seat, a foot the frame picked had
in the air read as down, its lock pulled the leg to the floor, and a
thigh turned 34.7 rad/s (the takes' own 22.1), a toe skating 6.7 mm.

The animator keeps the weights it gives the feet each frame, and a
state's transition (play) eases out of those. A matcher's change takes
its new frame's contacts as they are: its frames are re-picked every few
tenths of a second, and eased there a foot stayed locked as the frame
lifted it (32.2 rad/s at #906's check).

test_humanoid 22: on character.match_add over 100STYLE's neutral takes,
turned on the spot by 90 and 180 degrees (18.1 rad/s, the takes' own
22.1), then to a stool, sat on, stood up from and walked on: 15.2 rad/s,
no sole into the ground, a toe skating 3.7 mm. test_anim_pipeline's
matcher 17.8 rad/s within its 28.7. The motion suite's 25 programs pass.
motion.heading_off(m) reads how far a POWERED figure faces off the way
its animation holds its pelvis, what hold_heading holds.

test_handback_blend: the box man standing POWERED on its Walk, its
capsules resting on the floor. Its heading holds under 0.6 degrees while
it is upright; it then topples (tilted 0.49 rad by 0.5 s, falling from
about 0.6 s, lying from 2.2 s), and the heading #938 measured, 174
degrees in two seconds, was that fall. A grip torque at each stance foot
(tried) left it falling the same way. Printed, not held: the fall is
#968 (balance on a walk in place, #861's), and #938's heading is held to
once it stands.
…ree turn (#952)

character.use's walk to a slot and turn to its way skated a toe 7.4 mm
(test_use, the spot with no takes from 315 degrees). In the placement
turn's first step the body turned TURN_KEEP_OPEN, a quarter turn, past the
planted foot, which swivels nothing; twisted 90 degrees under the hips,
that foot was dragged as it lifted. TURN_KEEP_OPEN is now 1.1, a free
turn's TURN_REACH and TURN_PIVOT: the walk and turn skate 2.96 mm at the
most from all eight ways.

- test_use holds it: walking to a slot from any way and turning to its
  way, no toe skating 5 mm. 8/8.
- test_turn_spot reads a held foot at the point its lock holds (its ball,
  its toe's tip, or over its heel) rather than always at its ball. A
  placement turn's stance foot is held on its tip and rolls onto it as the
  body turns; its ball moved 22.7 mm while the tip held to 2.3 mm. 32/32.
… the own-heading lift (#1005, part)

The simulated feet behind mover.path_ways (feet_turn_sim_*) now:
- carry the placement turn (feet_turn_keep, the sim's 13th float): no
  pivot, TURN_KEEP_OPEN past a planted foot;
- lift the foot brought round on its own heading too once the pivot
  swivels, as the feet do (turn_step);
- path_ways keeps the facing from springing past the feet's bound, as
  mover.step does.

test_turn_spot reads mover.path_ways as each turn is asked and compares
its way with the figure's, frame by frame. Placement turns at 90 and 180
degrees each way come within 0.35 to 1.02 degrees and are held to 2
(36/36).

Not yet: a free turn's prediction is 6.3 to 9.8 degrees off at 90 and
180, all from round_outside, which the simulation does not model (with it
off in the feet, 0.5 degrees). Modelling it needs the pivot ball, hips
and stepping foot positions in the simulation.
The simulated feet behind mover.path_ways now carry where things stand
over the ground (ikrig's TURN_SIM floats, 31; the mover's array follows):
- the body's place, turned about the pivot's ball as pivot_move turns it;
- each foot's ball and ankle, the pivot's ankle swung round its ball as it
  swivels;
- the step under way's way round (round_pivot, from its ankle to its
  landing turned about what it goes round);
- the hips' middle and across off the body.
sim_round_outside bounds the simulated body as round_outside bounds the
figure's. Unmodelled, a free turn's way ahead ran 6.3 to 9.8 degrees
ahead of the figure's at 90 and 180 degrees.

test_turn_spot holds the mover's way ahead, read as each turn is asked,
within 2 degrees of the figure's, frame by frame: free turns 0.41 to 0.44
degrees, placement turns 0.35 to 1.02, 0 at 30. 42/42.

The cost of a simulated turn (a standing half turn, 240 steps ahead,
mover.path_ways): 10.9 us a call before, 26.9 after (45 to 112 ns a
simulated step).
… back edge (#977)

Two walks at the plaza steps' back side, the car plaza's worst joints
after the held-back release: into the face at the 0.45 m tread and held
there (seed 2's person 13, held at z -7.72), and up onto the 0.3 m tread
from behind and along its back edge, the 0.45 m one beside (seed 5's
person 16 at 3.40, 0.3, -7.47). Each reproduces its plaza event at the
same place: a calf and a foot at 60.8 rad/s and a toe thrown 7.76 m/s
held at z -7.72; a calf at 35.9 to 43.8 rad/s and a sole 17.7 mm into
the higher tread at x 3.40 on the 0.3 m one. Held to every check by
default; on today's code 4 of 77 checks fail, all four in these walks.
…own velocity, eased; a foot held back is let go as the body turns over it (#977)

mover.slid cuts the springs' velocity into the wall the capsule is held
back by to what the capsule went, the step it is held, and asks the
animator to ease the jump (animator.ease_params: the parameters' new
pose inertialized over WALL_EASE, 0.25 s, as a transition is), only for
a cut over WALL_CUT (0.1 m/s). "go" reads what is asked less what goes
into that wall (free_asked). Eased each frame for the few mm/s a slid
wall's way wavers by, a foot coming down skimmed the floor 12 mm
(test_feet_edges walk 10).

ikrig.lock_crossed: held back by a wall (feet_set_held), a locked foot
is let go once the hips come over it however the body goes: a body slid
along a wall turns to face the way it slides, and comes over the foot by
that turn. With the velocity cut, walk 10's walker turned to slide along
the steps, its left foot held on the 0.3 m tread came 8.2 mm past the
hips' middle; let go, 0.0.

test_character_walls, AE3D_WALLS_ALL=1, fastest joint / toe thrown,
8501e56a then now: swept square 54.4 / 2.18 then 22.1 / 2.33, into a
person 54.4 / 2.18 then 18.5 / 2.39, 27 degrees 32.2 / 2.87 then
23.8 / 3.97; default checks pass. test_feet_edges, AE3D_EDGES_ALL=1:
walk 19 (the steps' 0.3 m tread from behind) sole 17.7 to 0.0 mm, joint
43.8 to 22.9 rad/s; walk 14 joint 23.9 to 14.1, throw 2.79 to 0.93;
walk 18 throw 7.76 to 3.86 but joint 60.8 to 93.9; walk 15 joint 105.6
to 59.5 but sole 28.5 to 102.7 mm, a swing foot read the 0.45 m tread's
top as ground and went into the face below it; walk 10 throw 1.52 to
7.67, walk 11 1.46 to 2.29. Default checks: only walks 18 and 19 fail,
as before.
… that push as its velocity (#977)

slid cut the springs' velocity into the wall to what the capsule went
into it, and a capsule put back out of a neighbour's capsule went out
of it, negative: the cut set the velocity going away at the push's
speed. test_get_up_pair's first figure went 28 m across the floor before
it was knocked down, and the pair lay in other places, one layout 3.0 cm
inside each other getting up (the test failed). Cut now to no less than
standing against it.

test_get_up_pair: passes, the same 18 layouts as before the cut (waited
1.3 s at most). test_feet_edges AE3D_EDGES_ALL=1, walk 10 (along the
steps' open side into the deck's face): sole 2.34 to 0.0 mm, joint 29.0
to 20.5 rad/s, toe thrown 7.67 to 2.90 m/s; walk 15 into 103.5 mm,
joint 64.1, thrown 12.78; the rest the same. test_character_walls
AE3D_WALLS_ALL=1 the same but a person put inside its capsule (thrown
1.19 to 0.93).
…ide's room (#977)

The mover kept one way into what held its capsule back, the last one
blocked. In the corner the plaza deck's side makes with the steps' back,
a walker was asked along one face into the other, its velocity grew
along it and was cut each time the way switched: it stepped in place,
its feet 105 mm past the hips' middle and a calf at 69 rad/s (car plaza
seed 5's person 16, test_feet_edges walk 15). Now it keeps two faces
(m.wall, m.wall2, a held way within SAME_FACE of a kept one being that
face), the springs held back by each (held_by), and what is asked and
"go" go into neither (off_faces): wedged in a corner, nothing, and it
stands. The second face rides in the mover's state (its two spare
floats).

A face is let go once a move WALL_PROBE (now 0.1 m) into it is half
free. Tried 2 cm in, a walker stopped 1.5 cm short of the steps' back
face let go of it, started, came to it and stopped again, both feet
stepping onto the treads at the start, 44 cm in a frame (walk 18,
93.9 rad/s).

test_feet_edges AE3D_EDGES_ALL=1, 0885f2fa then now: walk 15 (the
deck's back corner) sole 103.5 to 0.0 mm, crossing 55.5 to 4.5 mm,
joint 64.1 to 23.6, toe thrown 12.78 to 1.33; walk 18 joint 93.9 to
58.3, skate 2.81 to 0.0; walk 11 throw 2.28 to 1.34; walk 10 joint 20.5
to 38.8 but throw 2.90 to 1.86; walk 19 throw 1.45 to 1.52. The rest
the same. test_character_walls AE3D_WALLS_ALL=1 the same.
…ds on (#977)

A foot planted on the road 1 mm from a kerb's face rolled its toe 2 mm
into the face at heel-off; the toe read the kerb's top as its ground,
the toes bent all they go and the foot was raised by what was left,
8 cm in a frame, a ball at 80 rad/s (test_feet_edges walk 6, steps 179,
182, 238). While a foot down has its toes read (sole_rest, toes_flat),
a toe point under a top whose face is within TOE_FACE (5 mm) of it
level (physics.static_way_out) reads the ground under the pad. Deeper
in, a foot on a stair with its toe in the next riser still stands on
that edge: with every toe ground over 6 cm above the pad's read as a
face, walk 9's foot went 41 mm into the riser.

test_feet_edges AE3D_EDGES_ALL=1: walk 6 joint 80.2 to 65.2 rad/s (sole
2.3 mm into the ground, under the 5 mm bound; its peak now step 97, a
calf), every other walk the same.
…977)

With "go" off where every face holds it (e9d60013), the facing still
went toward what was asked. Stopped by a neighbour's capsule and steered
round it, a walker stood and turned 2.7 rad/s toward the way asked on
its two locked feet, which crossed 290 mm (car plaza seed 5's person 23;
person 16 at 176 mm). Asked only into the faces holding it (free_asked
none), the facing is kept, and stand_turn reads it as standing. Going
along a face it faces the way asked, as before: turned to face along
it, a runner come into a block's side at 45 degrees turned a joint
18.0 rad/s, past test_character_walls' 17.6 bound (16.9 as before now).

Car plaza seed 5 (AE3D_TICK=60 FRAMES=STEPS=4800 EACH=1), e9d60013
then now: crossing 289.6 to 94.0 mm, joint 82.4 to 74.1 rad/s, toe
thrown 8.22 both, skate 16.7 to 16.8 mm, sole 142.6 to 80.0 mm; person
23 0 mm crossing (8501e56a: 26.9, 101.5, 5.00, 5.9, 80.0).
…ound on its top (#977)

A swing raised toward the plaza steps' 0.3 m tread (its look-ahead) was
let down when the ground ahead turned to the 0.45 m tread's face beside
it; its toe went 3 cm into the 0.3 m tread's face, sole_rest raised it
out of that "ground" 16 cm in a frame, and once off_walls put the toe out
of the face it dropped 25 cm in one, a calf at 58 rad/s
(test_feet_edges walk 18). A swinging foot's toe under a top it was
under last frame too (its toe as drawn then, now kept: toe_drawn, in the
feet's state) came to it through the face below it; unless the swing is
carried for that top this frame (foot_ground's ground, an edge ahead it
is raised for included), its toes read the ground under the pad there.
Without that last, toes swung at a kerb and a stair clip their risers
before the swing is over them and went 34 and 6 mm into them (walks 0
and 7).

test_feet_edges AE3D_EDGES_ALL=1: walk 18 joint 58.3 to 19.9 rad/s, its
peak now the swing raised for the 0.3 m tread and let down (steps 84 to
89); every other walk the same to the digit.
…rame, nor cuts one short (#977)

A walker held by the plaza steps' back face went into its stop on the
frame the mover cut its velocity into the face and asked the parameters'
jump eased (ease_params). The stop was taken and inertialized from what
was drawn, then eased again over the walls' quarter second: the trailing
swing came down over that and not the stop's own easing, and its toe was
swept 6.7 cm along the floor in a frame (3.86 m/s, test_feet_edges walk
18 at 1.72 s). The parameters' easing is now skipped on a frame a
transition is taken, and runs over no less than what is left of a
transition's.

test_feet_edges AE3D_EDGES_ALL=1: walk 18 thrown 3.86 to 2.06 m/s (its
peak now 1.57 s, a toe swung into the tread's face); every other walk
the same.
… edge half over a wall is not stepped up (#977)

Walk 18 (test_feet_edges), held at the plaza steps' back where the
0.3 m tread meets the 0.45 m one:
- off_walls read a solid as a step (not a wall) when it went no further
  than a step's height over the point in it. Over a toe a swing had
  carried 14 cm up, the 0.45 m tread was a step from the ground, the toe
  went 5 cm into its face, and was put out of it in a frame (2.06 m/s
  thrown at 1.57 s). The step's height is now over the ground the foot
  is on: its floor or the ground under it last frame, the higher.
- The look-ahead read the edge ahead under the toe alone: the swing was
  raised 21 cm for the 0.3 m tread under its toe while the 0.45 m one, a
  wall from the ground, was half under the foot, then let down as its toe
  came over that, a calf at 19.9 rad/s. Where a wall (wall_at: a top past
  ground_at's reach) is FOOT_HALF to either side of the way ahead, the
  lowest of the three reads is the ground ahead; elsewhere the toe's, as
  before (the lowest always, a swing across a kerb at 30 degrees was
  raised late, walk 1 at 23.9 rad/s, not 18.2).

AE3D_EDGES_ALL=1: walk 18 joint 19.9 to 15.3 rad/s, thrown 2.06 to 0.98,
sole 4.81 to 0.0 mm (every check now passes); every other walk the same
to the digit.
Beside the fastest joint's, the deepest sole's and the longest skate's
times: walk 18's 3.86 m/s throw was at 1.72 s, its stop, not where its
joint peaked.
… comes before it stops (#977)

foot_ground's look-ahead read the ground LOOK_AHEAD on along the toe's
way at its speed. A swing slowing to land short of the plaza steps'
0.3 m tread read the tread ahead at mid swing, was raised 15 cm for it,
then let down as it slowed and the point ahead fell short of the tread,
a calf at 22.5 rad/s (test_feet_edges walk 19, steps 56 to 61). The
point ahead is now no further than the toe comes to a stop at the rate
it slowed since last frame (its speed then: toe_speed, already in the
feet's state).

test_feet_edges AE3D_EDGES_ALL=1: walk 19 joint 22.5 to 19.5 rad/s (its
peak now the trailing foot let go a tread below its capsule, stepped up
the 0.3 m tread, steps 96 to 99); walk 7 16.1 to 14.1, walk 15 23.6 to
22.6; walk 0 23.1 to 27.5 and walk 6 65.2 to 66.9 (neither held to its
joint by default); every other walk the same.
…, under #1007 (#977)

Walk 19 (up onto the plaza steps' 0.3 m tread from behind) still turns a
calf 19.5 rad/s as its trailing foot, left on the ground a tread below
as the capsule stepped up, rises 30 cm right behind the riser (steps 96
to 99), and throws a toe 1.52 m/s carried 2.5 cm over the 0.15 m tread
at 2.65 s. Filed as #1007 (a swing past a riser it can step up is lifted
first, then carried forward). Its joint and throw checks move behind
AE3D_EDGES_ALL=1 with the header citing that; walk 18 keeps every check
by default, and no bound changes. Default run: all 75 checks pass.
A turn's step under way kept the landing it set off for: asked a half
turn and then back to 17 degrees (the car plaza, seed 4, a figure just up
setting off), the step went on round behind the body, crossed the feet
68 mm, and as the pivot let go its way jumped from round the ball to
straight and threw its toe 23 m/s. It now lands where the turn lands it
now (turn_land, from the heading it chose its landing from), and when its
landing or what it goes round changes, it goes on from where it is over
the rest of its travel, lifted as it was.

A turn step's heading as it was locked is read the way round the turn
goes: read the nearer way it flipped a whole turn as the body passed half
a turn from it, and a long step's toe swung 12 cm in a frame.

test_character_act: up and turned at once after a get-up, 90, -180, and
-180 asked back to 17 degrees mid-step: feet crossing under 5 mm, ending
within 2 degrees (the last crossed 89.6 mm without the change). Toes
printed, not held: the get-up's stance still gives turn steps 1.6 to 2.7
m/s and the foot it leaves off the ground slides up to 4 m/s.
…here it got to (#1006)

A turn's step set off again for a new landing kept its time: asked a
landing 4 rad round with 2% of its travel left, the rest went at 23 m/s;
asked one as it came down, it jumped 51 cm to it and threw its toe 29 m/s
(the car plaza, seed 4, a figure just up asked to walk off behind it).
Set off again, it now goes the rest of its way on a smoothstep of its own,
the whole step slowed where that way is longer than a turn's toe goes in
the travel left (the share done kept, so it is as high and as far along);
its travel done, it lands where it got to, the next step taking the rest
of the turn.
…t, steps timed for their way (#1006)

Right after a get-up's hand-back, toes went 2 to 4 m/s:

- The foot the get-up leaves off the ground was eased into the idle's
  stance along the floor, 13 cm at 1.5 to 2 cm up. At an act's end
  (character.act_stop) each foot not held within HOLD_NEAR of the ground
  is held down where it is (ikrig.feet_hold_down), against the clip's own
  contact, until the clip has it down, lifts it clear, or a step moves it:
  the stand's own step, lifted, puts it where the stand has it.
- A turn on the spot asked before an act was left on through it (the mover
  asks none while it acts): it stepped the turn's side foot as the get-up
  was handed back, and the next turn swung that foot half round, 95 mm
  past the hips' middle at 2.6 m/s. An act now turns it off, and so does
  its end.
- Timed for the turn alone, a step out of a get-up's wide stance (feet 62 cm
  apart) went round the pivot's ball at 2.8 m/s: a turn's step is timed for
  its ball's way round what it goes round too where that is longer
  (step_way; the simulated turn the same), and a foot put back under its
  stance as long as STEP_GAP_SPEED takes to cover its way.

Also: ikrig.feet_turn_keep is in the feet's state (a peer joining during a
placement turn turns it so), and character.keep_place sets it as an input.

test_character_act: up and turned 90, -180, and -180 asked back to 17
degrees, now holding no toe past 1.5 m/s (1.37, 1.45, 1.33 stepping; 0.74,
0.35, 0.17 not) besides the crossing and the facing.
Whether a figure's turns on the spot are placement turns (the hips kept
on their place, the feet stepped round them) is now in each tick's input
next to the facing asked (netanim.keep_place, I_KEEP, M_KEEP), so every
peer turns it as the owner does; a late join takes it from the snapshot's
inputs and the feet's state.

test_net_anim: a placement half turn added to the script after the turns
on the spot (every tick the same to the bit at no latency, and over 150 ms,
30 ms jitter and 5% loss), and a client joining in it, its planted feet
within 5 mm of the host's (0 mm), held; and again at no latency, its ticks
the same to the bit printed, not held (199 of 320: no late join is, #1009).
…1006)

Held down where a get-up left it (feet_hold_down), a foot 0.4 m off its
stance began a turn asked at once: it swivelled as the turn's pivot, the
body swung 0.7 m round its ball, and the walk asked set off late, its
pelvis then faster than its walk (test_balance seed 49, test_knockdowns
seed 37). As a person brings the far foot in before turning:

- A foot more than STEP_REACH off its stance before the turn has begun
  (no pivot swivelled, no turn's step under way) is put under it with a
  lifted step first (turn_step, off_stance, far_off); the turn's pivot
  does not swivel meanwhile, nor while a foot is held down, and the body
  waits for that step to land (feet_turn_reach).
- No turn step while a foot handed back is held down: the clip still eases
  into its stand, and a step to it chased it, a closing step's toe 1.9 m/s.
- A foot put back under its stance is timed for its way to the stance the
  step carries it to, and its ball's where that is longer.

A turn from a stance that is not wide is as before: test_turn_spot's
turns take the same steps and seconds (30 to 180 degrees free, one step;
placing, two).
…ist's range (#1010)

The showcase's rifleman (AE3D_WATCH=7), its hands solved onto the rifle
where the aim puts it, held its arms as a person never would: the elbows
bent as the idle bends them (hanging), so the hands hung off them and the
wrists bent 50 and 73 degrees onto the grips. Held straight, the wrists put
both elbows over the shoulders.

- ikrig.hands_hold_roll: a hold may let its hand roll round what it holds
  (its grip's axis, across the palm under the fingers) up to a given angle,
  as a hand takes a round grip at the angle its arm brings it; 0, the
  default, holds it as asked, as before.
- A hold that rolls puts its elbow as low as it goes with the wrist bent no
  more than WRIST_EASY (0.5 rad) off straight (the hand's rest turn in its
  forearm), found on the circle the arm's bones allow and the boundary
  halved (hold_elbow), the hand rolled a little a frame where that lowers
  it, and the elbow eased there (ELBOW_SETTLE); limb_solve_toward bends a
  limb toward a given middle point, less of it near the root-to-end line.
- The showcase's rifleman lets both hands roll 0.6 rad.

test_rifle_ready (new): the rifleman as the showcase sets him, his aim
swept: each elbow under its shoulder (13 and 17 cm), each wrist within
35 degrees (28 and 28; on b1437a1b 50 and 73), elbows bent no more than 150
(123), no shoulder raised over 5 cm (3.8), hands and elbows one frame with
the rifle (0.18 m/s at most in the chest's place).
…he ready (#1010)

A shouldered weapon's kind declares its butt and sight (rifle.item: an
M4's measures) and each grip's roll (a round fore-end rolls, a pistol grip
does not). Aimed, its butt goes in the trigger side's shoulder pocket,
taken from the pose each frame, the trunk turns its sight line onto the
point, and the eye on that side comes onto the line behind the sight
(animator.look_sight: the neck brings the eye, the head turns along the
line and stays upright). Ready states as a parameter (items.ready: aim,
high, low), eased at READY_HALF_LIFE: the muzzle 37 degrees over level or
32 under it, turned about a point between the butt and the shoulder's
joint, the head level.

Each hand is put on its own grip: the elbow as low as it goes with the
wrist within 0.5 rad, the hand rolled and slanted on a round grip as its
kind allows. The trigger elbow kept down by the ribs and never in across
the chest (ikrig.hands_hold_elbow_out; the owner chose elbows down on
2026-11-16). Items that declare nothing hold as before.

The showcase's rifleman is held through items and cycles aim, low ready
and high ready; AE3D_WATCH_ANGLE watches him from any side.

test_rifle_ready (all 12 checks), before this round / after:
  head cant aimed 35 / 0 degrees
  trigger elbow -31 to -19 / -9.4 to 0 degrees (bound -12)
  muzzle at the low ready -11 / -32 (bound -50 to -30), high 37 (30 to 45)
  butt off the pocket 0.04 cm aimed, 2.9 at a ready (bound 3)
  eye off the sight line 1.2 / 0.98 cm (2); sight off the point 0.17 deg (1)
  wrists at most 34.2 degrees (35); support elbow 18 cm under aimed (5)
  hands 0.002 mm off their grips; hands and elbows 0.14 m/s (0.5)
CI's linux-apps segfaulted at the showcase's exit: main freed the
characters first and the rifle after, and items.item_free lets go of its
holder's hands and aim (let_go_hands, aim_off), so it read the rifleman's
freed animator. MinGW's heap left the memory readable; glibc's did not
(the backtrace: main, item_free, let_go_hands). The items now go before
the figures, as the sword already did, and the rifleman's figure is kept
on the Show before its rifle is read, with kind_error printed should the
read fail, so a missing rifle cannot leave step_rifleman a null figure.
macOS's leak check on 2d4240b found the four Part measures and the list
holding them unfreed (5 leaks, 688 bytes).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant