Directed action with props and physics: methods and plan
Thu Nov 12, 23:45. This goes with game/ACTIONS.md, the action bible, which lists 55 gaps in 79 rows. It covers Don's directions of 23:10 and 23:20:
- Every action gets the right skeleton, clip and physics.
- Blender 5.2, on all three machines, is a first-class motion and physics tool.
- Masters are built at the highest quality, and quality is capped only at package time.
- No MetaHuman image or video ever goes to an AI model.
The Blender halves of the proof tests ran on Nov 13 ("Proof results" at the end). The Unreal halves have not run yet, so every method stays a recommendation until they do. The draft for studio METHODS is marked DRAFT.
1. The rule that decides most of this: the premiere replays the same
The film the player cuts plays at the premiere and must look the same every time. So nothing that simulates at runtime ships in a take. Physics is authored offline, in Blender or in the UE editor, baked, and played back as keys. That applies to rigid fracture, cloth, rope and tumbling props.
What the research found (Epic docs, forum and issue tracker, Nov 2026):
- Epic documents no determinism guarantee for live Chaos.
- Chaos cache playback has a run of bugs that appear only in packaged builds:
- UE-252163: the Cache Manager can't animate Nanite Geometry Collections in Static Pose mode.
- In 5.5, caches play in the editor but not in Shipping.
- In 5.6, the root proxy renderer breaks recording (UE-315559).
- 5.9 is the MegaJam preview, and its release notes flag longer cooks and some instability.
Keys are frame-indexed. They replay the same by construction, cost little, and survive packaging.
2. Best practice per interaction kind
Five things are judged for each kind: how a professional UE pipeline does it, what we can do this week with Blender 5.2, Kimodo on tower 1 and UE 5.9 on the laptop, whether it is deterministic, what it costs and the risk. Each kind ends with the recommended method and a 30-minute proof test.
Proof test rules:
- Blender parts run headless on the laptop CPU, so they need no GPU lock and can run while the other worker builds.
- UE parts wait for an editor slot on the laptop, or run in the tower 2 test editor. Check
NEWBOY\lab\.gpu.lockfirst. - Pass criteria are numbers read from bone and prop transforms, or ffmpeg measurements. Frames that show a MetaHuman never go to a model. Any visual check by a model uses a mannequin stand-in (Manny or a grey proxy on the same skeleton), as the self-critique rule already says.
2.1 Hand and foot contact (grips, plants, punches landing)
Pro pipeline. Contact is authored in the mocap or keyframe source, then fixed per shot:
- Control Rig in Sequencer: two-bone IK per limb, or Full Body IK when there are several contacts at once.
- Bake to an AnimSequence, or keep it as a live layered Control Rig track.
- Foot locking at plants.
- Motion Warping and Contextual Animation are gameplay tools (GASP traversal). They warp root motion to targets found at runtime and add nothing on a fixed timeline.
Our problem. The takes change clips after the fact with travel scales, keyed slide, lift and lift_keys, arc and tilt. Contact authored in clip space can't survive that. The prop transforms are known only in the take: stage_at and the set layouts.
Recommended method: world-space contact cleanup in Blender on a "world bake" of each take segment.
- Export each actor's evaluated pose per frame, with the actor transform and all modifiers applied, plus every prop's world transform. One action per actor per segment, from cue to cue.
- In Blender, put empties on the props' contact points: the rope grip, the cart top, the tailgate edge, the chair rail, the jaw.
- Add IK on the arms and legs with influence keyed on at the contact frames, and foot locks where a foot is planted. Read Kimodo's contact labels and the existing foot-slide measure (
game/motion/ab_measure.py) to find the plant frames. - Apply finger poses from the pose library (section 3).
- Bake with visual keying and export FBX on the same MetaHuman skeleton, with the root motion carried by the pelvis the way the aligned clips are now.
- The take then plays that clip with no
travel,slideorlift.
UE Control Rig in Sequencer stays for last-minute nudges only. Forum reports say Control Rig bakes come out offset in 5.8 and 5.8.1. Unverified on 5.9, so check each bake in the standalone AnimSequence view, not only in Sequencer.
- Deterministic: yes, it is a baked AnimSequence.
- Cost: about the same size as the clip it replaces (a few hundred KB, estimated), with no cook impact.
- Risk: the world-bake export is new tooling. The take builder's math has to be reproduced exactly, or the pose read back from a built sequence on tower 2 with the
retarget_export.pyread-back. Our own studies say IK keys land exactly (0.000 m wrist to target in brawl round 3), but whole-shot reads still failed on hand shape (butcher builds 1 and 2), so the fingers matter as much as the wrist.
Proof test P1 (30 min, Blender headless):
- Take rope_swing_kick_s1 as the MetaHuman FBX. Apply the take's 240 cm slide so this is the world-baked version.
- Add a rope grip empty at the take's staged rope point, with IK on both wrists for the hang frames and foot locks on takeoff and landing.
- Bake and export.
- Measure, from the exported FBX re-read in Blender: wrist to grip at most 1 cm on every hold frame, the planted foot slides at most 2 cm, and no joint jumps more than 4.5 cm in one frame.
- The UE half, when a slot is free: import onto MH-Rex, read back with
retarget_export.py, and check that the same numbers hold.
2.2 Prop attachment and hand-off (picking up the chair, holding a broken leg, the cap)
Pro pipeline. A socket on the hand bone plus the Sequencer Attach track, or Sequencer Constraints (Parent) keyed on at the grab frame. Compensation keys stop the pop, and the constraint is baked to keys before shipping.
Known traps:
- A Parent constraint fights a Transform track on the same actor (UE-238921).
- An Attach track loses its link when the subsequence asset moves.
Ours now. The chair is attached to hand_r for the whole take and hidden by scale keys, so it pops in and out (ACTIONS B10).
Recommended method:
- Short holds that end in a break (the chair): author the prop's motion in Blender in world space. A Child Of constraint keyed on at the grab frame and off at the release frame, then baked to world keys. The prop is one skinned asset with its own animation (section 2.3), so it needs no runtime attach at all.
- Long holds with nothing breaking (the cap, a megaphone, a cup): the socket plus the Attach track the builder already uses. The prop rests at its world spot until the grab frame and is attached on that frame with the relative transform it has on that frame (keep world), so it has no pop. Never use scale keys to hide a prop.
- Deterministic: yes.
- Cost: none.
- Risk: low if the attach frame is taken from the clip's grab frame, not guessed.
Proof test P2a (20 min, Blender):
- Put chair_0 at its beertent layout spot.
- Bake the heavy's world-baked brawl_chair_swing_s15.
- Key a Child Of to hand_r on the grab frame.
- Measure the chair's world position across the grab frame: under 0.5 cm a frame of jump.
- Measure both palms to the rail within 1.5 cm while held.
2.3 Breakables (the chair, the bandstand boards, the table, glass, crates, gate splinters)
Pro pipeline. Chaos Geometry Collections pre-fractured in Fracture Mode, broken by fields or strain, recorded with a Chaos Cache Manager and played from a Chaos Cache track. Or, very often in cinematics, the break is simulated offline and comes back as keys or as a pre-broken mesh swap behind a dust burst.
Recommended method: Blender Cell Fracture plus a rigid body sim, baked and exported as a skinned prop.
- Fracture the master mesh. Cell Fracture is an extension in Blender 4.2 and later; install it on all three machines, which the standing approval covers. Use 20 to 40 pieces for the chair and more for glass. Add hand-cut splinters for the wood.
- Simulate in Blender with the impact object animated as a kinematic collider: the table edge, the floor and Kid's back. Use the world-baked body from 2.1, so the pieces hit what the body actually hits.
- Bake with Rigid Body > Bake to Keyframes.
- Parent each fragment to its own bone, skin it rigidly (one bone, weight 1) and export FBX: a skeletal mesh plus one AnimSequence.
- In the take, the intact static mesh shows until the impact frame. The fractured skeletal mesh, in its rest pose (identical to the intact shape), takes over on that frame. Visibility keys only, so nothing pops.
Alembic geometry cache is the second-best export for rigid fracture. It stores every vertex every frame, where the skinned prop stores one transform per piece per frame. Chaos Geometry Collection plus a cache is the challenger, kept for a measured comparison only.
- Deterministic: the skinned prop yes. Alembic yes, because it is frame-indexed. The Chaos cache is not reliable in packaged builds (section 1).
- Cost, estimated until P2 measures it:
- Skinned prop: 30 bones x 90 frames compresses to tens of KB of animation, plus the mesh.
- Alembic: 20k vertices x 90 frames is about 21 MB raw before UE's codec.
- Chaos cache: grows with piece count x frames. No published figure.
- Risk: low for the skinned prop, because it uses the same import path as our characters. Alembic import needs the Alembic Importer and Geometry Cache plugins, and the Geometry Cache runtime has to cook. Chaos: high (section 1).
Proof test P2 (30 min Blender, then one UE slot):
- Fracture SM_breakaway_chair into 30 pieces.
- Simulate a drop onto a box edge and bake.
- Export (a) a skinned FBX and (b) an Alembic file.
- Record the file sizes.
- In UE, import both, put them in one test sequence, and play it twice in a Development package.
- Read back five piece transforms at three frames in each run. Pass: identical across runs (difference 0.0) and a uasset total under 2 MB for (a).
- Optional challenger: a Chaos Geometry Collection of the same chair with a cache. Same two-run read-back, and record the size.
2.4 Swinging and hanging (the rigging rope; tailgate and bed-edge hangs)
Pro pipeline. For cinematics: a skinned rope with an FK bone chain, keyed or simulated offline and baked. The character's hands are constrained to rope bones. The Cable Component is a cheap Verlet visual, depends on frame rate and can't carry a body cleanly. Physics Control and Control Rig Physics (beta in 5.8, more options in 5.9) are for live secondary motion, not a replay.
Recommended method: the rope as a bone chain in Blender, solved together with the character.
- Model the rope as a 16 to 24 bone chain from the tent anchor.
- Pin the bone at the grip point to the midpoint of Kid's two hands (his world-baked clip, with hand IK from 2.1) during the hold frames. Let the rest of the chain follow by a rigid body chain sim, or by Spline IK with a keyed pendulum.
- After release, a free pendulum that damps out.
- Bake both and export the rope as a skinned prop.
Hangs from an edge (C19 and K16) are the same idea without a rope: the hands are pinned to the edge empty by IK, and the body hangs from them. Never move the body with lift_keys under fixed hands.
- Deterministic: yes.
- Cost: small.
- Risk: the pendulum has to read as carrying an 80 kg man, so the swing period must match the rope length. A 3 m pendulum swings in about 3.5 s.
Proof test P3 (30 min, Blender):
- Build a 20-bone rope at the beertent rope spot.
- Drive it with the world-baked rope_swing_kick from P1 and bake.
- Measure the grip bone to the hand midpoint (at most 1 cm on hold frames) and the free-swing period after release (within 10 percent of 2π√(L/g)).
2.5 Rolling, tipping, swinging hinges and vehicles (cart, gate, truck, cannon recoil, crash pad)
Pro pipeline. Keyframes:
- A gate is one hinge rotation with eased overshoot and settle keys.
- A cart is its root translation plus wheel spin (distance over radius) plus a keyed tip curve.
- For a realistic tumble, simulate once in the editor and record it with Take Recorder, or simulate in a DCC and bake.
- Vehicles carrying a character: attach the character to the vehicle for the ride and author the clip in vehicle space.
Recommended method:
- Simple moves (gate swing, cannon recoil, truck suspension dip, crash pad squash): keyed in Blender or straight into the take JSON as move keys.
- Tumbles (the cart on the vault flub): a Blender rigid body sim baked to object keys.
- Wheels: the builder computes wheel rotation from the truck's keyed path.
- The tailgate drag (C11): parent Kid to the truck for the hang interval. The clip is world-baked in truck space, so his hands stay on the tailgate whatever the truck does.
- Deterministic: yes.
- Cost: none.
- Risk: a keyed tumble can look "animated". Bake the sim when it does.
Proof test P4 (30 min, Blender):
- Put the popcorn cart at its staged spot.
- Kid's world-baked vault_tumble_roll_on_s33 foot is a kinematic collider. Simulate the cart rocking or tipping, then bake to keys.
- Check: the foot contacts the cart top edge within 2 cm at the catch frame, and the cart comes to rest, below 1 cm a frame by the end.
2.6 Cloth and crowd (tent canvas, awnings, bunting, flags; drinkers and grandstand)
Pro pipeline.
- Cloth: Chaos Cloth (Cloth Asset, Panel Cloth Editor, Dataflow), recorded and cached for cinematics, or simulated in a DCC and brought in as an Alembic geometry cache. Background cloth is usually a material world-position-offset wind.
- Crowd: skeletal extras with offset clips up to a few dozen. Beyond that, AnimToTexture vertex animation on instanced meshes (City Sample), or Mass (far too heavy for the jam).
Recommended method.
Cloth:
- Hero cloth (the tent canvas at the rope anchor, an awning Kid brushes): Blender cloth bake exported as Alembic, imported as a Geometry Cache, played on a Geometry Cache track, a few seconds each.
- Background canvas, bunting and flags: a material wind whose time comes from a keyed parameter, not real time, so it replays the same.
- Costume jacket hem: Chaos Cloth stays render-only, as in
prep/CLOTH.md, but it is a known nondeterminism. P5 measures how far it varies between two runs. If it varies visibly, bake it the same way.
Crowd:
- Up to about 20 skeletal extras on the MetaHuman skeleton with offset Kimodo loops: cheer, clap, wince, laugh, drink.
- More than that: AnimToTexture with Nanite off, which is deterministic when its time input is keyed.
- Deterministic: yes for all of the recommended paths. Live Chaos cloth: no.
- Cost, estimated:
- A geometry cache of a 5k-vertex canvas over 4 s is a few MB after compression.
- Material wind: nothing.
- Each extra: a body, a costume and three or four clips.
- Risk:
- AnimToTexture breaks with Nanite and has reported material-node problems in 5.5.
- Extras are MetaHumans or not: if they use the MetaHuman skeleton, the 940 MB jam cap has to fit them. Use a low LOD and shared bodies.
Proof test P5 (30 min):
- In Blender, simulate a 5k-vertex canvas panel for 4 s with a wind gust, then export Alembic. Record the file size.
- In UE, import it as a Geometry Cache and play it twice in a Development package.
- Read back the vertex positions of 3 vertices at 3 frames through Blueprint or Python on the component. Pass: identical.
- Record the uasset size.
- In the same package, run the Chaos jacket hem twice and record its largest vertex difference.
2.7 Seams between clips and the rejoin after a miss
Pro pipeline.
- In Sequencer: overlapping sections with blend-in and blend-out weights, plus the motion blending tools on each section ("Match With This Bone In Previous Clip", root offsets, root motion trail).
- In gameplay: inertialization and Dead Blending in the AnimGraph, and motion matching (Pose Search) picking poses from a database.
Those three gameplay tools need an AnimBP that owns the pose. Our sequencer owns Rex's pose (chase.take.json notes), and motion matching is wasted on a fixed script. The Animation Mixer (5.8, experimental, on the also-experimental UAF) has an open Dead Blending crash.
STANDARDS (Game director, line 23) bars root sliding to rejoin a path: "the branch's own motion ends where the main path continues."
Recommended method: continuity authored into the motion, not repaired at the seam.
- Generate in continuity. A branch clip is generated by Kimodo with its first frame constrained to the main clip's pose at the commit cutoff, and its last frame constrained to the main path's pose and root at
rejoin_t. Kimodo takes root and joint constraints (research/motion.md). - World-bake each branch from the branch point to the rejoin frame, as one continuous clip in Blender. Close any residual gap there by spreading it over the whole branch with root and IK correction, never as a ramp at the end, and lock the feet while doing it.
- Crossfades: keep the 0.25 s Sequencer crossfade with "Match With This Bone In Previous Clip" (pelvis), only for main-path clip-to-clip joins that already share a pose.
A build check measures every seam: pelvis position at most 1 cm and pelvis velocity within 10 percent across the join, feet locked on planted frames, and no ramp in the root keys. This goes with the parallel seam audit (game/takes/SEAMS.md when it lands) and the flawless-branches ticket.
- Deterministic: yes.
- Cost: one more clip per branch.
- Risk: Kimodo constraint quality on the end pose is unproven. If it drifts, Blender closes the gap.
Proof test P6 (30 min):
- For the brawl's early_sell flub, world-bake the main brawl_beat_a_s17 and the branch haymaker_early_sell_s2.
- Close the gap to the main path at
rejoin_tin Blender: a pelvis correction spread over the branch, with foot locks. - Measure the root key curve for a ramp (no velocity step over 10 percent), the seam pelvis error (at most 1 cm) and the foot slide on plants (at most 2 cm).
2.8 Keyed moves Kimodo can't make (the cannon flight, the wig pull-on, fine hand business)
Recommended method: key them by hand in Blender.
- Use Auto-Rig Pro, already in the studio, or a light IK rig built on the exported MetaHuman skeleton with the same bone names.
- Key the poses for the flight: launch stretch, arms forward, the slight roll in the air, the tuck. The pelvis follows the take's existing parabola (
arc), imported as a path, so the timing stays on the beats. - Export onto MH-Rex.
The wig pull-on is in no take yet. It belongs to the opening or a makeup-chair beat (STORY: Kid wears Rex's costume and wig). It needs a wig prop with a hand-off (2.2) and finger poses (section 3).
- Deterministic: yes.
- Cost: about 1 to 2 hours of keying per move.
- Risk: the self-critique has to be numeric or done on a stand-in body.
2.9 Masters and package profiles
Don, 23:11: masters at the highest quality, caps only at package time.
Masters:
- Blender
.blendsources with full piece counts and full-rate bakes at 30 fps, or 60 where a fast break needs it. - Uncompressed FBX and Alembic, kept outside the UE project. On tower 1, outputs follow the backup rules.
Two package profiles, set in config only:
- Jam profile, under 940 MB:
- Device profiles and texture LOD bias.
- Animation compression (ACL) settings.
- Static mesh LOD and Nanite fallback settings.
- Geometry cache compression level, and which cloth caches are included.
- A reduced-piece fracture variant generated by script from the master, not by editing it.
- Ultra profile: the same project with the caps lifted, for the post-jam Director's Cut edition.
Rule: no source asset is ever overwritten with a lower-quality version.
3. Hands and fingers
- Kimodo's generator drives about 30 joints. The SOMA77 export has hand joints, but the fingers carry no authored motion.
- Our retarget A/B read-back carries only the middle metacarpal and finger 1.
- The MetaHuman body has full finger chains, so every grip in the game is currently whatever the retarget leaves in the fingers.
- Kimodo's SOMA hands are not usable for grips. We need a pose library.
Pose library (Blender asset library, applied as a finger-only layer and keyed in and out over 3 to 4 frames):
| Pose | Used by | Shape |
|---|---|---|
| grip_rope | B14, P3 | Power grip around a 3 to 4 cm cylinder, thumb wrapped |
| grip_rail | B10 (chair back and leg), K07 (stair rail) | Power grip around a 2.5 to 3 cm rail |
| hook_edge | C11, C12, C18, C19, K16 (tailgate, bed edge) | Fingers curled 90 degrees over an edge, thumb alongside |
| palm_plant | C05, C06 (cart top), K09 (muzzle rim), C09 (mirror push) | Flat palm, fingers spread and slightly extended |
| fist | B03, B17, B19 (punches) | Closed fist, thumb outside |
| high_five | K04, K05 | Open palm, fingers together |
| key_press | S01 (squawk box) | Index extended, other fingers curled |
| relaxed | default for every clip | Natural half-curl, replaces the retarget rest pose |
| pinch | the wig pull-on, the cap | Thumb and index pinch |
Then a contact pass in the manner of brawl-grip round 5 and grip_solve2.py: curl each finger until it touches the prop's surface, then stop. Fingertips within 1.5 cm of the surface on at least 80 percent of hold frames.
4. Plan: laptop tickets in priority order
Each ticket lists its method, its proof test and the motion still needed. "Kimodo" means a batch on tower 1. Every prompt goes through the motion gate and gag checks as now.
- T1 World-bake tooling and the Blender contact pipeline. Export each take segment's evaluated per-frame world poses (all actors) and prop transforms to a
.blend. Re-import cleaned clips onto MH-Rex, MH-Heavy and MH-Walt. Takes then droptravel,slideandliftfor cleaned segments.- Method: 2.1. Proof: P1.
- Motion needed: none.
- Blocks T3 to T8.
- T2 Finger pose library. The nine poses in section 3 on the MetaHuman hand, plus the contact curl.
- Method: section 3. Proof: fingertip distance on P1's rope grip.
- Motion needed: keyed poses only.
- T3 Brawl choreography. Fixes B03, B06, B08, B15 and B17.
- Re-sequence the heavy: heavy_wild_swing_miss_s12 for the b1 duck, heavy_haymaker_throw_s10 into the b3 sell.
- Add a shove for b4.
- Contacts: fist to jaw, feet to chest, back on the table top.
- Remove the 30 cm uppercut slide.
- Method: 2.1. Proof: P1 numbers on the fist to jaw (at most 2 cm at the hit frame).
- Kimodo: "a big man steps in and shoves another man hard in the chest with both hands, arms fully extending", 3 s, 8 seeds.
- T4 The breakaway chair, and the table's end. Fixes B10 and B07, and note 1.
- Pick-up from chair_0, a two-hand hold, a fracture on the table edge, the leg kept then dropped, pieces left on the floor.
- Optionally the trestle folds after the slide (beat 2's line).
- Method: 2.2 and 2.3. Proof: P2a and P2.
- Kimodo, for the eat_the_chair wreck: "a big man swings a wooden chair two-handed across another man's back, follows through and stumbles a step", 3 s.
- T5 Rope swing. Fixes B14, B16 and B25.
- A bone-chain rope, hands pinned to it, the 240 cm keyed slide removed (the body's travel comes from the swing), the tent canvas ripple at the anchor.
- Method: 2.4 and 2.6. Proof: P3 and P5.
- Motion needed: none new. The rope_swing_kick_s1 world bake carries it.
- T6 Truck and tailgate. Fixes C11, C12, C18, C19, C21 and C22.
- Kid is parented to the truck for the drag and the hangs; hook_edge hands; the bed height is set from the real flatbed mesh; lift pops are replaced by a pull-up clip that ends on the bed.
- Wheels spin from the path, the body pitches on start and stop, and the truck reverses for the wreck reset.
- Method: 2.5, 2.4 and 2.1. Proof: P4, plus hands to edge at most 1 cm while dragged.
- Kimodo: "a man hanging from a ledge by both hands swings a knee up onto it and climbs up to stand", 3.5 s.
- T7 Cannon. Fixes K07, K09, K11, K12, K13 and K14.
- Steps or a ladder on the platform with a keyed climb.
- The climb-in with palm_plant on the rim and the body going into the bore; the cannon hides the body from the waist down.
- A keyed Blender flight (2.8).
- Recoil, flash and smoke on fire.
- Truck suspension dip on landing and pad squash on the flop.
- Method: 2.8, 2.5 and 2.1. Proof: the P1 numbers on the rim and step contacts; for the flight, the pelvis on the arc at most 1 cm.
- Kimodo: "a man runs up five steep wooden steps two at a time and stops on top, one hand on the rail", 2.5 s, with root height constraints at each step.
- T8 Miss variations for every HUD-only cue. Fixes C13, C20, B05, B18, K05, K10 and K15, as STANDARDS requires. This is where the flawless-branches ticket and the Fable trial apply.
- Method: 2.7. Proof: P6 on each.
- Reuse: heavy_shrug_off_s11 for b2 and s1.
- Kimodo, each with start and end constraints from the main path:
- "a man hanging from the back of a moving truck lets go too early, drops, stumbles three steps and keeps running" (c5, t2)
- "a man sitting on the rim of a large tube scrambles, slips and drops inside late" (c1_in)
- "a man in mid air opens out of a tuck too early, lands stiff-legged and stumbles forward two steps" (c3)
- For MH-Walt: "a man holds up his right hand for a high five, waits, looks at his hand, lowers it slowly and shakes his head" (k2)
- T9 Bandstand fall. Fixes B19 and W05.
- Move the bandstand to the fall spot, or the fall to the bandstand (the sets worker owns it).
- Raise Kid to the deck height before the punch; breakaway deck boards and rail; a tray of glasses goes over (B24).
- Method: 2.3. Proof: P2 on a board panel.
- Motion needed: none new.
- T10 Chase props. Fixes C05, C06, C07, C08, C09, C10, C14, C16 and C17.
- The real popcorn cart with palm_plant on the vault and a rock or tip on the flub.
- A keyed gate swing closing over the slide, and the bump and reset.
- Mirror panels on Kid's path that flex on the push.
- A hay pile at the landing that compresses and scatters.
- Method: 2.5, 2.3 and 2.1. Proof: P4.
- Kimodo, for MH crew: "a stagehand walks to a gate, pulls it back open and steps away", 3 s.
- T11 Money man and crowds. Fixes C03, B23 and K19.
- Cast the money man (a MetaHuman or a shared body) with a run clip.
- 12 to 20 extras in the beer tent and 20 or more seated in the grandstand.
- Method: 2.6.
- Kimodo:
- "a heavyset man runs hard, glancing back over his shoulder twice"
- "a man standing at a bar cheers and pumps his fist"
- "a seated spectator claps and cheers, then leans forward"
- "a man winces and turns his head away"
- "a man laughs and slaps his knee"
- Loops of 4 s, gated for a matching start and end pose.
- T12 Beats and cloth dressing. Fixes S01, S02, K20, B25 and S05.
- Walt reaches and presses the squawk box key (key_press).
- Rex does craft-table business: cup, mirror, hair, keyed or Kimodo.
- Material wind on the flags and bunting.
- Kimodo, for MH-Walt: "a man leans down to a small box on a low stand, presses a button with his index finger and talks into it, then straightens up".
- Kimodo, for MH-Rex: "a vain man checks his hair in a hand mirror, picks up a coffee cup, sips, and puts it down", 4 s.
- T13 Seam check in the build. The section 2.7 numbers on every boundary and every rejoin. Fails the build on a ramp. Shared with the seam audit.
- T14 Walk-through consistency. Fixes W01, W05 and W06: the same platform and rope assets as the takes, and Walt offset from Kid's step.
- T15 Kung fu take (Director's Cut). Fixes B26. Write
kungfu.take.jsononce T3 proves the fight pipeline.
Proof tests P1 to P6 come before T3 to T12, in this order: P1 (it gates everything), P2a and P2, P3, P4, P6, P5. P1 to P4 and P6 need only Blender on the laptop CPU. The UE halves wait for an editor slot.
Batch 17 picks (brawl and cannon motion, Thu Nov 12, 23:37 to 23:40 on tower 1)
Job mj-kimodo-actions-17, Kimodo SOMA-RP v1.1, prompts and durations from the "Kimodo:" lines of T3, T4 and T7 (T1, T5 and T9 list no new Kimodo motion). Script game/motion/kimodo_actions17.py, gag rules in game/motion/gag_checks.py. Output and ranking.md are in ~/studio-ai/renders/megajam/mj-kimodo-actions-17/ on tower 1 (previews of the top 3 per move in previews/). No Kimodo constraints were used: the existing batch scripts have no constraint path and the plan's start and end poses need the main-path clips as input, so the first and last frames are not locked to the main path and the step heights are judged by the gag rule alone.
| move | ticket | takes | passing | best tag | note |
|---|---|---|---|---|---|
| brawl_shove_chest | T3 | 8 | 0 | brawl_shove_chest-s4 (gag FAIL) | Gag: both wrists drive forward faster than 1.8 m/s together at chest height. The best takes reach only about 1.5 to 1.8 m/s, so the shoves are soft. Rank 1 is the best joint-limit score (0 violations), not a pass. |
| brawl_chair_swing_2h | T4 | 20 | 0 | brawl_chair_swing_2h-s4 (gag FAIL) | Gag: both wrists together above the shoulders for 6 frames, then a fast swing. Only s4 gets both hands up at all (4 frames, then a 3.1 m/s swing); 19 takes swing from the side or below the shoulder. |
| cannon_climb_steps | T7 | 20 | 9 | cannon_climb_steps-s20 | Gag: feet alternate at least 4 times and the pelvis rises at least 0.5 m in steps. Top 3: s20, s8, s14. Joint-limit score is high (155 for the best), so expect a cleanup pass. |
Zero passing takes: the shove and the two-handed chair swing, the two brawl moves. Next step for them is a rerun with more seeds and a stronger prompt (for example "lunges forward and drives both palms into his chest"; "lifts a wooden chair above his head with both hands and slams it down and across"), or a Blender keyed shove from the P1 pipeline.
Proof results: the Blender halves (Fri Nov 13, 00:10, tower 2)
Run headless in Blender 5.2.1 on tower 2 (CPU only, no GPU lock needed), from D:\studio-ai\blender_test\actions_proofs. No Unreal editor was opened and nothing was rendered. Every number below is read from bone and object transforms, and the body numbers come from the exported FBX after it was imported back into a clean scene.
Setup. Each Kimodo BVH was retargeted onto the MetaHuman body skeleton (D:\megajam\retarget\ab\rig\MH_body.fbx, exported from SKM_Body) with the A/B challenger script, and every clean-up was exported on that skeleton. Scripts are in game/proofs/scripts/ and the per-test result JSON in game/proofs/results/.
Three findings that change the methods:
- Blender's IK constraint fails on the imported UE skeleton. No bone is connected, the importer guesses the bone tails, and the bone axes are flipped on one side. A wrist aimed 10 cm away ended 30 cm off. Every contact below therefore uses a scripted analytic two-bone IK in world space (
two_bone_ikinplib.py), keyed per frame. It lands exactly: 0.000 cm on every pinned wrist and foot. - Rigid Body > Bake to Keyframes needs a UI context and fails headless. The script does the same job: it keys each piece from the point cache, and the keys match the cache to 0.0 cm.
- The exported skeleton's root bone is
pelvis. It has norootbone. The UE half has to check that the import maps onto the MH skeleton. If it won't, add arootbone at the origin before export.
| Test | Result | Key numbers | Exported for the Unreal half |
|---|---|---|---|
| P1 contact, rope_swing_kick-s1 | Partial: contacts pass, pops fail | Wrists to grip 0.000 cm. Fingertips to rope surface: median 0.57 cm, max 0.64 cm, 100% within 1.5 cm. Planted foot (ankle) slide 0.0 to 1.98 cm; ball 2.75 to 2.95 cm at toe-off. Joint acceleration 41.1 cm/frame² against 15.9 in the source. | D:\megajam\proofs\actions-2026-11-13\p1_rope_swing_kick_s1_contact_1.fbx (4.36 MB) |
| P1b truck-space contact, truck_hang_on-s37 | Pass | Feet on the bed: 0.004 cm of slide in total while the same feet travel 11.06 m in the world. Left hand to rail at most 0.375 cm while the truck accelerates, pitches 3° and heaves 4 cm. Re-read foot to lock 0.0 cm. | D:\megajam\proofs\actions-2026-11-13\p1b_truck_hang_on_s37_truckspace.fbx (4.36 MB) |
| P-finger, the 9-pose library | Pass | Curl-to-contact grip_rope: each fingertip's bone line 0.21 to 0.64 cm from the surface (the pad is 0.8 cm). grip_rail 0.51 to 0.76 cm. Pinch tip gap 1.36 to 1.38 cm. FBX round trip 0.0 cm. | game/proofs/finger_lib.fbx (2.24 MB, one pose per frame), game/proofs/results/finger_lib.json |
| P2a hold, brawl_chair_swing-s15 | Pass with a two-hand hold. The plan's method, Child Of on hand_r alone, fails | With Child Of hand_r alone, the left palm ends up to 107 cm off the rail. With the two-hand hold: palms to rail 0.22 and 0.00 cm, and the chair jumps 0.43 cm at the grab with an 8-frame keep-world ease (2.90 cm with no ease). | D:\megajam\proofs\actions-2026-11-13\p2a_heavy_chair_swing_s15_hold.fbx (4.64 MB), game/proofs/p2a_chair_hold_prop.fbx (0.88 MB), game/proofs/p2a_chair_world_keys.json |
| P2 breakable chair | Pass | 30 pieces from the 19,998-triangle chair in 4.7 s, with 70 glue pairs. Sim 0.72 s for 90 frames. The chair stands still before the hit (0.21 cm) and shatters on the hit (30 of 30 pieces move more than 30 cm). Two full runs (fracture plus sim) are identical (0.0 cm, 0.0 rotation). Skinned FBX 1.64 MB, Alembic 2.53 MB, intact mesh 0.83 MB. Re-read pieces 0.0 cm. | game/proofs/p2_chair_fracture_skinned.fbx, game/proofs/p2_chair_fracture.abc, game/proofs/p2_chair_intact.fbx |
| P3 rope, 20 bones | Pass | Grip on P1's hands: pinned segment end at most 0.78 cm, the next bone's head 1.56 cm (chain stretch). Free swing period 3.045 s against 3.026 s for a hanging chain. With the 75 kg proxy: hands to grip 0.14 cm, loaded period 3.875 s against 3.883 s (double pendulum) and 3.844 s (2π√(L/g)). After release the rope swings at 3.061 s. Both runs identical. | game/proofs/p3_rope_body75_swing.fbx (0.94 MB), game/proofs/p3_rope_grip_release.fbx (0.77 MB) |
| P4 cart, vault_tumble_roll_on-s33 | Fail | Palms on the counter 0.0 cm, but only after lifting the body up to 59.5 cm: the clip is a dive roll at ground level, not a vault. The run-up path goes through the cart, so the feet strike its base at about frame 61. The feet are kinematic, with infinite mass, so they throw the 70 kg cart 6.6 m and tip it 172°. The cart comes to rest on its side (0.0 cm/frame), and two runs are identical. | Not worth importing. D:\megajam\proofs\actions-2026-11-13\p4_vault_tumble_s33_palms.fbx is kept for the record |
| P5 cloth (Blender half) | Pass | Canvas: 5,041 vertices, 3 x 2.5 m, top edge pinned, keyed gust, 120 frames. Bake 24.0 s, two runs identical, billows 28.6 cm, Alembic 35.9 MB (7.3 MB of raw positions: Blender also writes normals and UVs every frame), re-read 0.0006 cm. Bunting: 12 pennants, 504 vertices, bake 1.44 s, identical runs, 16.6 cm of movement, Alembic 2.85 MB. | D:\megajam\proofs\actions-2026-11-13\p5_canvas.abc, game/proofs/p5_bunting.abc |
| P6 seam, haymaker_early_sell-s2 into brawl_beat_a-s17 | Partial: velocity fails, the rest pass | Pelvis at the branch point 0.0 cm and at the rejoin 0.0 cm. The raw gap was 17.6 cm, plus 3.45 cm of height at the start. The correction moves at most 0.354 cm/frame and 0.0095 cm/frame at both ends, so there is no ramp. Pelvis acceleration 1.13 cm/frame² against 1.14 raw. Plant slide 0.0 cm (raw up to 4.9). Pelvis speed at the branch point: main 1.08 cm/frame, branch 0.21 (an 86% mismatch). | D:\megajam\proofs\actions-2026-11-13\p6_haymaker_early_sell_s2_rejoin.fbx (3.74 MB) |
| P-keyed cannon flight | Pass | Pelvis to the take's arc at most 0.0017 cm over 108 frames. 6 key poses, from the launch stretch to the tuck, reached within 0.002 cm. The largest joint acceleration is 10.0 cm/frame², at film 29.58 s, where the take's tilt keys go from 0 to 60° in 0.15 s. | D:\megajam\proofs\actions-2026-11-13\p7_cannon_flight_keyed.fbx (4.38 MB) |
What the failures need:
- P1 pops. The source hold is only 7 frames (59 to 65) with the hands 1 m apart, so pinning both hands to one rope moves the arms up to 31 cm in a frame. The method works, but the clip doesn't fit it. Fix: a rope clip from Kimodo with the hands held together through a longer hang, or the hang keyed in Blender (2.8). The take's own 240 cm slide drags the planted feet 137 to 160 cm. Moving the same travel into the airborne frames (60 to 75) is what made the feet pass.
- P4. It needs a real vault clip, for example Kimodo "vaults over a waist-high cart, planting both hands on top", with the cart staged off the run-up. The flub's foot should be a dynamic collider with a body's mass, or the tip keyed from a gentler sim. The tip, the rest and two identical runs already work; the contact staging is what failed.
- P6 velocity. A position correction spread over the branch adds no velocity at either end, so it cannot fix a speed mismatch the clips already have. The branch has to be generated in continuity, by Kimodo with start and end constraints. Batch 17 had no constraint path. Also, the branch is 75 frames but rejoin_t is 103 frames after the cue, so it holds its last pose for 0.93 s.
- P2a. Hold two-handed props in a frame built from both palms, both pinned. Child Of on one hand works only for one-handed props.
Settings that matter:
- The rope rigid body runs at 60 substeps and 80 solver iterations. At 200 and 200 the loaded period went wrong (1.35 s).
- Cloth mass is per vertex: 0.002 kg for the 5k canvas, which is 10 kg in total.
- Cell fracture without the extension is a bmesh Voronoi bisect (
p2_break.py). Glue thresholds from 25 to 1000 made no difference against a 9 m/s fist, so T4 tunes the glue against the real impact.
Recommended order for the Unreal half:
- P2 chair: the skinned FBX against the Alembic, played twice in a Development package with the pieces read back and the uasset sizes noted. It is the lowest risk and gates T4 and T9.
- The skeleton check, folded into the first body import onto MH-Rex with the
pelvisroot, then read back withretarget_export.py. - P1b truck-space clip attached to the truck (T6).
- P3 rope, skinned (T5).
- P-keyed flight (T7).
- P2a heavy with the chair prop (T4).
- P5 Geometry Cache plus the Chaos hem variance. It carries the most UE risk, with two plugins to cook.
- P6 after a constrained Kimodo rerun.
- P1 and P4 after new motion.
Batch 17b picks (rerun of the two brawl moves, Fri Nov 13, 00:19 to 00:24 on tower 1)
Job mj-kimodo-actions-17b (it waited from 23:41 behind a sound-effects job and batch 16b), script game/motion/kimodo_actions17b.py, 40 seeds each, same gag rules and ranking. Prompts: the shove is "a big man lunges forward and shoves another man violently in the chest with both hands at once, driving both arms fully straight out fast with a strong push, then follows through leaning forward"; the chair swing is "a big man holds a heavy wooden chair above his head with both hands, then swings it down and across another man's back with a huge fast two-handed swing, follows through and stumbles a step". Output and ranking.md are in ~/studio-ai/renders/megajam/mj-kimodo-actions-17b/ on tower 1.
| move | takes | passing | best tag | note |
|---|---|---|---|---|
| brawl_shove_chest | 40 | 5 | brawl_shove_chest-s3 | Passing: s3, s26, s2, s29, s33. s3 and s26 also have 0 joint-limit violations. Batch 17 had 0 of 8. |
| brawl_chair_swing_2h | 40 | 0 | brawl_chair_swing_2h-s22 (gag FAIL, rank 1 on joint limits only) | Still zero pass. Only 5 takes raise both hands over the shoulders; closest is s40 (8 frames up, 3.2 m/s swing against the 3.5 needed), then s13 (10 frames, 2.1 m/s). |
Fallback reference for the chair, batch 8 brawl_chair_swing (40 takes, mj-kimodo-brawl-8): 22 of 40 pass the older gag (a wrist peak of 4.5 m/s or more, any grip), with peaks up to 7.5 m/s (s11, s21, s10, s14, s36). None pass the two-handed overhead rule. They are fast chair swings, not overhead starts, so they work for the swing across the back but not for a raise. The two-handed chair swing is to be keyed in Blender (the T4 / P2a path), using batch 8 s11 or batch 17b s40 as timing reference.
Batch 18: constrained generation and the P1, P4 and P6 clips (Fri Nov 13, tower 1)
Jobs mj-kimodo-actions-18 (01:15 to 01:21, 346 s) and mj-kimodo-actions-18b (01:23 to 01:32, 518 s), Kimodo-SOMA-RP-v1.1 only. Code: game/motion/kimodo_constrained.py (the constraint path), kimodo_actions18.py, new gag rules vault_over and rope_hang in gag_checks.py, seam table in rank_takes.py. Results and ranking.md are on tower 1 in ~/studio-ai/renders/megajam/mj-kimodo-actions-18 and ...-18b.
The constraint path (P6 result: passes). Kimodo's FullBodyConstraintSet is fed the first and last 3 frames (3 frames pin the pelvis velocity as well as the pose). The poses come from the main path read out of brawl.take.json: start = brawl_beat_a-s17 at film 5.6733 s (clip time 1.157 s, rate 0.7), end = brawl_beat_b-s1 (from 8.73 s, travel 0.5) at the rejoin. The clip is 104 frames, so it reaches the rejoin at 9.1067 s, 0.01 s after rejoin_t (set rejoin_t to 9.1067 in the build). The branch is now as long as the gap, so it no longer holds its last pose. B is placed on A by continuity at the join (a 0.31 degree turn), because there is no engine offline; check it against a build.
| P6 seam, 40 seeds of haymaker_early_sell_c2 | Start (commit) | End (rejoin) | Target |
|---|---|---|---|
| Pose gap, max joint (cm) | 0.01 | 0.02 | under 2 |
| Pelvis offset (cm) | 0.0 | 0.0 | under 2 |
| Pelvis velocity mismatch | 0% (main 0.395 cm/frame) | 0% (main 0.662 cm/frame) | under 20% |
| Joint acceleration in the 9 frames around the seam, new vs source (cm/frame²) | 1.55 vs 1.55 | 1.55 vs 1.55 | match |
All 20 seeds of round 1 and all 40 of round 2 meet both targets. The ends are pinned, so the seam figures say the plumbing works; what a seed can still get wrong is the middle (gag, joint limits, peak acceleration: the whole-clip peak is 4.6 to 15.9 cm/frame² for the gag passers against 3.4 for the source, so prefer the low ones).
Round 1 failed its gag, round 2 passed. Round 1 (the plan's prompt, default guidance): 0 of 20 passed lying_travel; the body never went below 0.80 m. Round 2 (a stronger prompt, text guidance 4.0): 22 of 40 pass the gag and the seams.
Batch 18 picks
| Ticket | Pick | Backups | Gag pass | Notes |
|---|---|---|---|---|
| P6 early_sell branch | haymaker_early_sell_c2-s25 | s28 (seam accel 2.51 at one end), s37, s24, s11 | 22/40, seams 40/40 | Joint-limit violations 4 (s28: 1, s24: 2). Not yet viewed by eye. |
| P4 cart vault | none that I trust. cart_vault_side_b-s15 passes the gag | s38, s22 fail it | 1/40 (round 1: 0/30) | s15 has 68 joint-limit violations over 56 frames: unusable until looked at. Plants were 10 frames at 1.2 m/s, pelvis 0.11 m over the hands. |
| P1 rope swing | none | closest: round 2 s34, 6 frames of hands within 15 cm (0.2 s); round 1 s21, 10 frames within 25 cm | 0/70 | Needs 0.6 s (18 frames). Kimodo does not hold both hands together on a rope from text alone. |
What P1 and P4 need next. The two clips whose contact is the whole gag are the two Kimodo cannot do from text. Options: (1) constrain them too, with hand-position end-effector constraints on a rope or cart point over the hang or the plant frames (the EndEffectorConstraintSet class is installed; the new path only does full-body so far); (2) key the hang in Blender (section 2.8) from the P3 rope; (3) for the cart, take the P4 vault_over frames from s24 or s26 (pelvis 0.26 to 0.31 m over the hands at 3.5 m/s) and clean the legs in Blender. I stopped after two rounds, as the ticket rules say.