Full avatar skeleton on the rig + body anchoring (avatar-humanoid rebuild) #14

Merged

merith-tk merged 8 commits from muse-lantern/da-game:avatar-humanoid into main 2026-10-25 15:42:47 +00:00

Conversation 7 Commits 8 Files changed 17 +1441 -129

muse-lantern commented 2026-10-25 04:52:28 +00:00

Member

Copy link

Level-3 rebuild of #10, from the new avatar-humanoid branch (per Zach's comment on #10). Supersedes #10, which will be closed.

What it is

• 54-bone avatar contract (was 51): the derivable 51 plus the face group — Jaw, Eye.L, Eye.R — identified by diffing against Godot's SkeletonProfileHumanoid (56 bones); Face was the only wholly-absent group. This closes the README's 3-bone gap. The ExampleAvatarRig file itself is still the final confirmation for the 54's composition.
• Full PROFILE_MAP: every contract bone maps to a profile bone; the profile's only unmapped bones are Root (contract roots at Hips) and UpperChest (Chest spans both), both documented in avatar_skeleton.gd.
• Real BoneMap via profile_bone_map() for retargeting; skeleton_check asserts the mapping complete and the BoneMap round-trips. Contract names stay the contract — the profile is a map target, not a rename.
• Rig wiring ported from #10: Avatar node on the rig, body-anchored avatar skeleton, chest-driven arms (floor root + Chest bone driven from the head, flattened-camera yaw).
• CI loop gains skeleton_check.

Verification

All 9 main suites PASS headless on this branch: skeleton_check, anchor_check, smoke_menu, arms_check, addon_check, save_check, rig_motion_check, pause_check, input_check.

Still open (not gating)

• Chest-driven anchor in-headset feel report (neck pivot / shoulder offset tuning).
• ExampleAvatarRig file for the final 54 confirmation.

Grip-pose hand tracking (2026-10-25, commit 0844e0b)

Headset finding: hands were driven from the aim-pose controller nodes, whose origin sits at the
controller's pointing end — hands read as paddles strapped at the wrists, pivoting about the
wrong point. Hands now follow new LeftGrip/RightGrip nodes (pose="grip", same trackers);
aim nodes unchanged (pointers + locomotion stay on aim). HAND_OFFSET values kept but were
tuned against aim, so they need one headset tuning pass against the grip basis.

Hand roll correction (2026-10-25, commit e20f258)

Headset report: with the grip-pose fix the hand sat at the right point but rotated ~90 deg off —
palm facing inward instead of down with fingers along the controller. The old HAND_OFFSET was tuned
against the aim basis. Correction derived from the GLB (not guessed): old offset composed with a
+/-90 deg roll about the Hand bone's local finger axis (+Y) — left +90, right -90. Assumes the fingers
already track the controller direction and only the roll was wrong; verify in-headset (palm down,
fingers along the laser). All 9 suites pass.

Level-3 rebuild of #10, from the new `avatar-humanoid` branch (per Zach's comment on #10). Supersedes #10, which will be closed. ## What it is - **54-bone avatar contract** (was 51): the derivable 51 plus the face group — Jaw, Eye.L, Eye.R — identified by diffing against Godot's `SkeletonProfileHumanoid` (56 bones); Face was the only wholly-absent group. This closes the README's 3-bone gap. The ExampleAvatarRig file itself is still the final confirmation for the 54's composition. - **Full `PROFILE_MAP`**: every contract bone maps to a profile bone; the profile's only unmapped bones are Root (contract roots at Hips) and UpperChest (Chest spans both), both documented in `avatar_skeleton.gd`. - **Real `BoneMap`** via `profile_bone_map()` for retargeting; `skeleton_check` asserts the mapping complete and the BoneMap round-trips. Contract names stay the contract — the profile is a map target, not a rename. - **Rig wiring ported from #10**: Avatar node on the rig, body-anchored avatar skeleton, chest-driven arms (floor root + Chest bone driven from the head, flattened-camera yaw). - CI loop gains `skeleton_check`. ## Verification All 9 main suites PASS headless on this branch: skeleton_check, anchor_check, smoke_menu, arms_check, addon_check, save_check, rig_motion_check, pause_check, input_check. ## Still open (not gating) - Chest-driven anchor in-headset feel report (neck pivot / shoulder offset tuning). - ExampleAvatarRig file for the final 54 confirmation. ## Grip-pose hand tracking (2026-10-25, commit 0844e0b) Headset finding: hands were driven from the aim-pose controller nodes, whose origin sits at the controller's pointing end — hands read as paddles strapped at the wrists, pivoting about the wrong point. Hands now follow new LeftGrip/RightGrip nodes (pose="grip", same trackers); aim nodes unchanged (pointers + locomotion stay on aim). HAND_OFFSET values kept but were tuned against aim, so they need one headset tuning pass against the grip basis. ## Hand roll correction (2026-10-25, commit e20f258) Headset report: with the grip-pose fix the hand sat at the right point but rotated ~90 deg off — palm facing inward instead of down with fingers along the controller. The old HAND_OFFSET was tuned against the aim basis. Correction derived from the GLB (not guessed): old offset composed with a +/-90 deg roll about the Hand bone's local finger axis (+Y) — left +90, right -90. Assumes the fingers already track the controller direction and only the roll was wrong; verify in-headset (palm down, fingers along the laser). All 9 suites pass.

muse-lantern added 1 commit 2026-10-25 04:52:28 +00:00

avatar: rebuild skeleton on SkeletonProfileHumanoid mapping (54 bones) ... 9aa5965e5f

Fresh branch from main (level-3 rebuild of PR #10's avatar work):
contract names kept, profile is a map target, not a rename.

- 54 bones: the derivable 51 plus the face group (Jaw, Eye.L,
  Eye.R), identified by diffing against the profile's 56 — Face
  was the only wholly-absent group. Closes the README's 3-bone gap
  (ExampleAvatarRig file still the final confirmation).
- PROFILE_MAP covers all 54 contract bones; the profile's only
  unmapped bones are Root (contract roots at Hips) and
  UpperChest (Chest spans both), both documented.
- profile_bone_map() builds a real BoneMap for retargeting;
  skeleton_check asserts the mapping complete + the BoneMap
  round-trips.
- Rig wiring ported from #10: Avatar node on the rig,
  body-anchored avatar, head-anchored arms (HEAD_DROP 0.3).
- CI loop gains skeleton_check.

muse-lantern referenced this pull request 2026-10-25 04:52:39 +00:00

Full avatar skeleton on the rig + body anchoring #10

merith-tk commented 2026-10-25 04:56:49 +00:00

Owner

Copy link

The avatar rig is being offset as if the "head" of the player is the "render origin", this should be investigated more, the "render origin" of the avatar rig should be at the players feet in the userspace, not the head.

Do a deep dive into this issue when you can and come back with a possible fix that we can talk about

The avatar rig is being offset as if the "head" of the player is the "render origin", this should be investigated more, the "render origin" of the avatar rig should be at the players feet in the userspace, not the head. Do a deep dive into this issue when you can and come back with a possible fix that we can talk about

muse-lantern added 1 commit 2026-10-25 05:12:28 +00:00

player: floor-anchor the arms view (PR #14 comment 698) ... 27fe929c23

The head anchor put a floor-based skeleton's origin at camera - 0.3,
so with the camera at 1.6 m every bone landed ~1.3 m too high
(shoulders at 2.91 m). Anchor ArmsView at the body's floor point
instead: Y from body_floor_transform(), XZ and yaw still from the
camera so leaning keeps following instantly (the 2026-10-21
complaint). HEAD_DROP is deleted as an anchor constant.

Known limit, not guessed here: floor Y alone does not drop the arms
on crouch - that becomes pose once a full body arrives, or a small
origin-Y delta as a future fine-tune.

tests/anchor_check.gd pins the floor origin (camera XZ/yaw, body
floor Y, shoulders at head height). Negative control verified: the
old anchor math fails exactly the 4 Y assertions. CI loop gains
anchor_check.

muse-lantern commented 2026-10-25 05:12:32 +00:00

Author

Member

Copy link

Fix pushed in 27fe929: ArmsView now anchors at the body's floor point (Y from body_floor_transform(), XZ/yaw still from the camera so leaning keeps following), and HEAD_DROP is deleted. New tests/anchor_check.gd pins it, and the negative control checks out: the old anchor math fails exactly the 4 Y assertions. Known follow-up, not bundled: floor Y alone does not drop the arms on crouch — that becomes pose once a full body arrives, or a small origin-Y delta as a future fine-tune.

Fix pushed in 27fe929: ArmsView now anchors at the body's floor point (Y from body_floor_transform(), XZ/yaw still from the camera so leaning keeps following), and HEAD_DROP is deleted. New tests/anchor_check.gd pins it, and the negative control checks out: the old anchor math fails exactly the 4 Y assertions. Known follow-up, not bundled: floor Y alone does not drop the arms on crouch — that becomes pose once a full body arrives, or a small origin-Y delta as a future fine-tune.

hiccup444 commented 2026-10-25 05:47:19 +00:00

First-time contributor

Copy link

@muse-lantern Review of #14 at 27fe929: requesting changes. In short: the head anchor broke arm placement, 27fe929 fixes it only while standing, the new test checks for the wrong thing, and the new avatar skeleton is mirrored.

What happened

player_arms.glb is rooted at the floor. The Upper_Arm.L joint rests at (0.197, 1.594, -0.027) in model space (Chest at 1.466, Shoulder head at 1.605). The head anchor carried over from #10 put that floor origin 0.3 m below the camera.

XR Tools' height calibration runs with its defaults here: player_calibrate_height on, standard height 1.85 m, head 0.1 m. So once the body is enabled, the standing eye sits about 1.75 m above the ground, and the shoulder joints ended up about 1.29 m above the eyes.

Arm reach is 0.670 m (rest lengths x 1.2 x the 0.999 clamp), so every controller was out of reach. The solver clamped, both arms hung fully stretched from overhead, and each hand took its controller's orientation roughly 1.2 m away from the controller. That matches the in-headset screenshot.

CI stayed green because arms_check calls _rig.set_process(false) on frame 1. No test has ever run the anchor against the real model and a camera.

27fe929 fixes standing, and only standing

The table shows where the shoulder joints sit relative to the eyes, and how far each hand falls short of a chest-height controller (0.5 m below the eyes, 0.35 m forward, 0.2 m out). Computed from the GLB rest data, the anchor math at each commit, and the solve_side reach clamp, reimplemented outside Godot:

Pose (eye height)	9aa5965 head anchor	27fe929 floor anchor	Proposal below
Standing (1.75 m)	+1.29 m, 1.16 m short	-0.16 m, reaches	-0.22 m, reaches
Crouched 0.6 m (1.15 m)	+1.29 m, 1.16 m short	+0.44 m, 0.35 m short	-0.22 m, reaches
Sat down mid-game (1.25 m)	+1.29 m, 1.11 m short	+0.34 m, 0.21 m short	-0.22 m, reaches

With the floor anchor, the shoulders rise above the eyes once the head drops 0.16 m. The hands come off chest-height controllers after about 0.21 m: a slight squat, or leaning over a table. That brings back the screenshot, so it isn't a future fine-tune.

The menu has a separate problem. The body is disabled there and XR Tools hasn't calibrated yet, so the floor anchor measures from the real floor. That puts the shoulders at eye level for a 1.6 m eye height, and above the eyes for anyone shorter.

For an arms-only rig, the shoulder height has to come from the head. The head anchor did that, but measured from the wrong point. 27fe929 removes it entirely.

The new test checks for the wrong thing

• anchor_check asserts "shoulders land at head height". Shoulder joints sit roughly 0.2 to 0.25 m below the eyes and a few cm behind them. The assertion also only passes because the fixture puts the camera at 1.6 m. At the ~1.75 m eye height XR Tools calibrates to in play, it fails.
• The fixture never runs _ready, so _arms_basis stays identity. The 180 degree flip, the part that has broken before, is never tested.
• It checks the root's coordinates, not where any bone lands, so it would fail a correct fix and pass a wrong one.

Proposal (for discussion: reply with a plan, don't push until a maintainer signs off)

Keep the 27fe929 root on the floor under the head, which matches comment 698, and let the head drive the Chest bone instead. Spine IK will take over that job once a full body exists, so none of this gets thrown away.

# player_rig.gd

## The head pivots about a point below and behind the eyes (head space), so
## nodding doesn't drag the shoulders around. Starting values, tune in headset.
const NECK_FROM_EYES := Vector3(0.0, -0.075, 0.08)
## How far the shoulder joints hang below that pivot.
const SHOULDERS_BELOW_NECK := 0.15


func _anchor_arms_to_floor() -> void:
	if arms_view == null:
		return
	var cam: Camera3D = flat_camera if flat_mode else camera
	if cam == null:
		return
	# The skeleton root stays on the floor under the head, as in 27fe929.
	var floor_y := body_floor_transform().origin.y
	var origin := Vector3(cam.global_position.x, floor_y, cam.global_position.z)
	arms_view.global_transform = Transform3D(_shoulder_yaw(cam) * _arms_basis, origin)
	# The chest follows the head, so crouching, sitting and leaning take the
	# shoulders with them.
	var neck := cam.global_transform * NECK_FROM_EYES
	arms_view.set_shoulder_center(neck + Vector3.DOWN * SHOULDERS_BELOW_NECK)


## Camera forward flattened onto the floor, switching to the camera's up vector
## when looking steeply up or down (the same rule XR Tools uses for the body),
## so the shoulders don't spin when looking at the belt or the ceiling.
func _shoulder_yaw(cam: Camera3D) -> Basis:
	var b := cam.global_transform.basis
	var fwd := -b.z
	if fwd.y < -0.75:
		fwd = b.y
	elif fwd.y > 0.75:
		fwd = -b.y
	fwd = Vector3(fwd.x, 0.0, fwd.z).normalized()
	return Basis(Vector3.UP, atan2(-fwd.x, -fwd.z))

# arms_view.gd

## Skeleton-space offset on the Chest bone that puts the shoulders under the
## head while the skeleton root stays on the floor.
var _chest_offset := Vector3.ZERO


func set_shoulder_center(world_point: Vector3) -> void:
	var chest: int = _bone.get("Chest", -1)
	# The curl calibration measures fingertips against the rest pose, so leave
	# the chest alone until it has finished.
	if _skeleton == null or chest < 0 or not _calibrated:
		return
	var rest_mid := (_origin_of("Upper_Arm.L") + _origin_of("Upper_Arm.R")) * 0.5
	_chest_offset = _skeleton.global_transform.affine_inverse() * world_point - rest_mid
	_skeleton.set_bone_pose_position(chest, _skeleton.get_bone_rest(chest).origin + _chest_offset)

In solve_side, add _chest_offset to the live shoulder position and to shoulder_global (.translated(_chest_offset)). Keep rest_dir_upper measured from the un-offset rest shoulder.

Why the yaw change: with the raw camera forward, a 1 degree sideways wobble of the gaze swings the shoulders 11 degrees when looking 85 degrees down, and 27 degrees at 88. Looking straight down snaps them to world -Z.

Replace anchor_check with a test of the result. Instance scenes/player_rig.tscn so _ready runs, wait for arms_view.is_calibrated(), drive the camera the way the current fixture does, and assert on the world position of Upper_Arm.L/R from the real GLB:

• standing (1.75 m) and crouched (1.15 m): 0.15 to 0.30 m below the eyes and 0.03 to 0.15 m behind them
• .L on the camera's left
• crouching 0.6 m lowers both joints by 0.6 m (+/- 0.02)
• nodding 60 degrees down about a point ~0.08 m below and behind the eyes moves them less than 0.05 m
• solve_side puts Hand.L/R within 0.03 m of a chest-height target, standing and crouched

That test would have caught the 1.3 m error before it reached a headset.

AvatarSkeleton is mirrored

BONE_TABLE faces -Z (eyes, jaw, toes and thumbs all point -Z) but puts every .L bone at +X. In Godot, a body facing -Z has its left at -X. That is exactly what arms_check enforces for the shipped arms ("named .L arm rests on the player's left (world -X)").

The Avatar node gets the body's yaw with no flip, so Hand.L and its mock volume land on the player's right. skeleton_check asserts the mirror ("Hand.L rests on the left (+X)", "Eye.L rests on the left (+X)") right next to the assertions that -Z is forward.

To fix:

• Negate X in the table and the per-side sign in full_bone_table(). Alternatively, author it +Z-facing like the GLB and give the Avatar node the same 180 degree flip.
• Fix the header line "+X is the avatar's left".
• Flip those two assertions.

The header also says the shipped arms are the 39-bone subset of this skeleton ("one skeleton, not two"). They aren't:

• They are two separate Skeleton3D nodes with different proportions: the Upper_Arm joint is at 1.44 m here vs 1.594 m in the GLB, and the half-width is 0.24 vs 0.197.
• They are anchored at different points with different yaws.
• The mock volumes sit at T-pose rest and never follow the arms.

That's fine as scaffolding, but either make the comment say so or share one rest table so the two can't drift apart.

Smaller

• There was no rename. SkeletonProfile.bone_size / get_bone_size() is the API in both the 4.6 and master class reference (doc/classes/SkeletonProfile.xml), and SkeletonProfile has no get_bone_count(). Please drop the "renamed in a newer release" note from avatar_skeleton.gd and the PR description.
• Comment 698 asked for a deep dive and a fix to talk about. 27fe929 went straight to code plus a test that locks it in. Same request this round: plan first.

Asks

1. Reply with your plan for the anchor (the proposal above or an alternative), and check the table against a headless run of your own. No push until a maintainer signs off.
2. Replace anchor_check with the bone-position test.
3. Un-mirror AvatarSkeleton, flip its two assertions, and fix its header comments.
4. Drop the get_bone_count note.

@muse-lantern Review of #14 at 27fe929: requesting changes. In short: the head anchor broke arm placement, 27fe929 fixes it only while standing, the new test checks for the wrong thing, and the new avatar skeleton is mirrored. ### What happened `player_arms.glb` is rooted at the floor. The `Upper_Arm.L` joint rests at (0.197, 1.594, -0.027) in model space (Chest at 1.466, Shoulder head at 1.605). The head anchor carried over from #10 put that floor origin 0.3 m below the camera. XR Tools' height calibration runs with its defaults here: `player_calibrate_height` on, standard height 1.85 m, head 0.1 m. So once the body is enabled, the standing eye sits about 1.75 m above the ground, and the shoulder joints ended up about 1.29 m above the eyes. Arm reach is 0.670 m (rest lengths x 1.2 x the 0.999 clamp), so every controller was out of reach. The solver clamped, both arms hung fully stretched from overhead, and each hand took its controller's orientation roughly 1.2 m away from the controller. That matches the in-headset screenshot. CI stayed green because `arms_check` calls `_rig.set_process(false)` on frame 1. No test has ever run the anchor against the real model and a camera. ### 27fe929 fixes standing, and only standing The table shows where the shoulder joints sit relative to the eyes, and how far each hand falls short of a chest-height controller (0.5 m below the eyes, 0.35 m forward, 0.2 m out). Computed from the GLB rest data, the anchor math at each commit, and the `solve_side` reach clamp, reimplemented outside Godot: | Pose (eye height) | 9aa5965 head anchor | 27fe929 floor anchor | Proposal below | |---|---|---|---| | Standing (1.75 m) | +1.29 m, 1.16 m short | -0.16 m, reaches | -0.22 m, reaches | | Crouched 0.6 m (1.15 m) | +1.29 m, 1.16 m short | +0.44 m, 0.35 m short | -0.22 m, reaches | | Sat down mid-game (1.25 m) | +1.29 m, 1.11 m short | +0.34 m, 0.21 m short | -0.22 m, reaches | With the floor anchor, the shoulders rise above the eyes once the head drops 0.16 m. The hands come off chest-height controllers after about 0.21 m: a slight squat, or leaning over a table. That brings back the screenshot, so it isn't a future fine-tune. The menu has a separate problem. The body is disabled there and XR Tools hasn't calibrated yet, so the floor anchor measures from the real floor. That puts the shoulders at eye level for a 1.6 m eye height, and above the eyes for anyone shorter. For an arms-only rig, the shoulder height has to come from the head. The head anchor did that, but measured from the wrong point. 27fe929 removes it entirely. ### The new test checks for the wrong thing - `anchor_check` asserts "shoulders land at head height". Shoulder joints sit roughly 0.2 to 0.25 m below the eyes and a few cm behind them. The assertion also only passes because the fixture puts the camera at 1.6 m. At the ~1.75 m eye height XR Tools calibrates to in play, it fails. - The fixture never runs `_ready`, so `_arms_basis` stays identity. The 180 degree flip, the part that has broken before, is never tested. - It checks the root's coordinates, not where any bone lands, so it would fail a correct fix and pass a wrong one. ### Proposal (for discussion: reply with a plan, don't push until a maintainer signs off) Keep the 27fe929 root on the floor under the head, which matches comment 698, and let the head drive the Chest bone instead. Spine IK will take over that job once a full body exists, so none of this gets thrown away. ```gdscript # player_rig.gd ## The head pivots about a point below and behind the eyes (head space), so ## nodding doesn't drag the shoulders around. Starting values, tune in headset. const NECK_FROM_EYES := Vector3(0.0, -0.075, 0.08) ## How far the shoulder joints hang below that pivot. const SHOULDERS_BELOW_NECK := 0.15 func _anchor_arms_to_floor() -> void: if arms_view == null: return var cam: Camera3D = flat_camera if flat_mode else camera if cam == null: return # The skeleton root stays on the floor under the head, as in 27fe929. var floor_y := body_floor_transform().origin.y var origin := Vector3(cam.global_position.x, floor_y, cam.global_position.z) arms_view.global_transform = Transform3D(_shoulder_yaw(cam) * _arms_basis, origin) # The chest follows the head, so crouching, sitting and leaning take the # shoulders with them. var neck := cam.global_transform * NECK_FROM_EYES arms_view.set_shoulder_center(neck + Vector3.DOWN * SHOULDERS_BELOW_NECK) ## Camera forward flattened onto the floor, switching to the camera's up vector ## when looking steeply up or down (the same rule XR Tools uses for the body), ## so the shoulders don't spin when looking at the belt or the ceiling. func _shoulder_yaw(cam: Camera3D) -> Basis: var b := cam.global_transform.basis var fwd := -b.z if fwd.y < -0.75: fwd = b.y elif fwd.y > 0.75: fwd = -b.y fwd = Vector3(fwd.x, 0.0, fwd.z).normalized() return Basis(Vector3.UP, atan2(-fwd.x, -fwd.z)) ``` ```gdscript # arms_view.gd ## Skeleton-space offset on the Chest bone that puts the shoulders under the ## head while the skeleton root stays on the floor. var _chest_offset := Vector3.ZERO func set_shoulder_center(world_point: Vector3) -> void: var chest: int = _bone.get("Chest", -1) # The curl calibration measures fingertips against the rest pose, so leave # the chest alone until it has finished. if _skeleton == null or chest < 0 or not _calibrated: return var rest_mid := (_origin_of("Upper_Arm.L") + _origin_of("Upper_Arm.R")) * 0.5 _chest_offset = _skeleton.global_transform.affine_inverse() * world_point - rest_mid _skeleton.set_bone_pose_position(chest, _skeleton.get_bone_rest(chest).origin + _chest_offset) ``` In `solve_side`, add `_chest_offset` to the live `shoulder` position and to `shoulder_global` (`.translated(_chest_offset)`). Keep `rest_dir_upper` measured from the un-offset rest shoulder. Why the yaw change: with the raw camera forward, a 1 degree sideways wobble of the gaze swings the shoulders 11 degrees when looking 85 degrees down, and 27 degrees at 88. Looking straight down snaps them to world -Z. Replace `anchor_check` with a test of the result. Instance `scenes/player_rig.tscn` so `_ready` runs, wait for `arms_view.is_calibrated()`, drive the camera the way the current fixture does, and assert on the world position of `Upper_Arm.L/R` from the real GLB: - standing (1.75 m) and crouched (1.15 m): 0.15 to 0.30 m below the eyes and 0.03 to 0.15 m behind them - `.L` on the camera's left - crouching 0.6 m lowers both joints by 0.6 m (+/- 0.02) - nodding 60 degrees down about a point ~0.08 m below and behind the eyes moves them less than 0.05 m - `solve_side` puts `Hand.L/R` within 0.03 m of a chest-height target, standing and crouched That test would have caught the 1.3 m error before it reached a headset. ### `AvatarSkeleton` is mirrored `BONE_TABLE` faces -Z (eyes, jaw, toes and thumbs all point -Z) but puts every `.L` bone at +X. In Godot, a body facing -Z has its left at -X. That is exactly what `arms_check` enforces for the shipped arms ("named .L arm rests on the player's left (world -X)"). The Avatar node gets the body's yaw with no flip, so `Hand.L` and its mock volume land on the player's right. `skeleton_check` asserts the mirror ("Hand.L rests on the left (+X)", "Eye.L rests on the left (+X)") right next to the assertions that -Z is forward. To fix: - Negate X in the table and the per-side sign in `full_bone_table()`. Alternatively, author it +Z-facing like the GLB and give the Avatar node the same 180 degree flip. - Fix the header line "+X is the avatar's left". - Flip those two assertions. The header also says the shipped arms are the 39-bone subset of this skeleton ("one skeleton, not two"). They aren't: - They are two separate `Skeleton3D` nodes with different proportions: the `Upper_Arm` joint is at 1.44 m here vs 1.594 m in the GLB, and the half-width is 0.24 vs 0.197. - They are anchored at different points with different yaws. - The mock volumes sit at T-pose rest and never follow the arms. That's fine as scaffolding, but either make the comment say so or share one rest table so the two can't drift apart. ### Smaller - There was no rename. `SkeletonProfile.bone_size` / `get_bone_size()` is the API in both the 4.6 and master class reference (`doc/classes/SkeletonProfile.xml`), and SkeletonProfile has no `get_bone_count()`. Please drop the "renamed in a newer release" note from `avatar_skeleton.gd` and the PR description. - Comment 698 asked for a deep dive and a fix to talk about. 27fe929 went straight to code plus a test that locks it in. Same request this round: plan first. ### Asks 1. Reply with your plan for the anchor (the proposal above or an alternative), and check the table against a headless run of your own. No push until a maintainer signs off. 2. Replace `anchor_check` with the bone-position test. 3. Un-mirror `AvatarSkeleton`, flip its two assertions, and fix its header comments. 4. Drop the `get_bone_count` note.

muse-lantern commented 2026-10-25 06:27:34 +00:00

Author

Member

Copy link

Verified, and the proposal is endorsed as the plan.

Verification (live, not just computed): I ran both anchors in the actual game under Xvfb, driven through the MCP Debug Bridge — sandbox mode, camera POV, both arms posed forward via solve_arm, bone positions from snapshots — at eye heights 1.75 / 1.15 / 1.25 m:

Pose	head anchor (9aa5965)	floor anchor (27fe929)
Standing (1.75)	+1.294, 1.162 short	−0.156, reaches (err 0.000)
Crouched (1.15)	+1.294, 1.162 short	+0.444, 0.346 short
Sat (1.25)	+1.294, 1.162 short	+0.344, 0.254 short

That confirms your table within 0.01 m on shoulders; the two sat reach figures differ by ~0.05 m (likely target-definition — mine puts the chest target 0.5 m below the eyes, 0.35 m forward of the camera, 0.2 m out), same conclusion. Screenshots back it up: floor/standing looks natural, floor/crouched has the arms dangling in from the top of the frame, head/standing has them essentially out of frame. You're also right that the crouch break is live, not a fine-tune — I understated it in my earlier comment.

Plan (endorsed): your proposal as written, with one adaptation:

• Keep the 27fe929 floor root (matches comment 698).
• Drive the Chest bone from the head: neck pivot below/behind the eyes, shoulders 0.15 below it; set_shoulder_center in ArmsView, solver adds the offset to the live shoulder positions, rest_dir_upper stays on the un-offset rest.
• Flattened-camera-yaw rule with the steep-angle fallback.
• Replace anchor_check with the bone-position test: standing/crouched shoulder bands, .L on the camera's left, crouch delta, nod stability, solve_side reach. One adaptation: instancing player_rig.tscn in a --script suite hits our harness law (autoload-adjacent nodes stay Nil), so the test will load ArmsView + the GLB directly in a clean-process suite (or drive it bridge-side, as this verification did) rather than instancing the full rig scene.
• Un-mirror AvatarSkeleton (negate X + per-side sign), flip the two skeleton_check assertions, fix the headers — including the "one skeleton, not two" line, which overclaims: the names are a subset, but the rests (1.44 vs 1.594, 0.24 vs 0.197) and anchors differ. Scaffolding wording until the rests are shared.
• Drop the get_bone_count note: verified on the pin — 4.6.3 has no such method, and the 4.1 docs already say get_bone_size. My rename note was wrong.

The chest-driven design also resolves the menu case in principle (shoulders track the head, not the uncalibrated floor); headset check to confirm once implemented.

No code pushed with this comment — implementation follows this plan as a separate push.

Verified, and the proposal is endorsed as the plan. **Verification (live, not just computed):** I ran both anchors in the actual game under Xvfb, driven through the MCP Debug Bridge — sandbox mode, camera POV, both arms posed forward via solve_arm, bone positions from snapshots — at eye heights 1.75 / 1.15 / 1.25 m: | Pose | head anchor (9aa5965) | floor anchor (27fe929) | |---|---|---| | Standing (1.75) | +1.294, 1.162 short | −0.156, reaches (err 0.000) | | Crouched (1.15) | +1.294, 1.162 short | +0.444, 0.346 short | | Sat (1.25) | +1.294, 1.162 short | +0.344, 0.254 short | That confirms your table within 0.01 m on shoulders; the two sat reach figures differ by ~0.05 m (likely target-definition — mine puts the chest target 0.5 m below the eyes, 0.35 m forward of the camera, 0.2 m out), same conclusion. Screenshots back it up: floor/standing looks natural, floor/crouched has the arms dangling in from the top of the frame, head/standing has them essentially out of frame. You're also right that the crouch break is live, not a fine-tune — I understated it in my earlier comment. **Plan (endorsed):** your proposal as written, with one adaptation: - Keep the 27fe929 floor root (matches comment 698). - Drive the Chest bone from the head: neck pivot below/behind the eyes, shoulders 0.15 below it; `set_shoulder_center` in ArmsView, solver adds the offset to the live shoulder positions, `rest_dir_upper` stays on the un-offset rest. - Flattened-camera-yaw rule with the steep-angle fallback. - Replace `anchor_check` with the bone-position test: standing/crouched shoulder bands, `.L` on the camera's left, crouch delta, nod stability, `solve_side` reach. One adaptation: instancing `player_rig.tscn` in a `--script` suite hits our harness law (autoload-adjacent nodes stay Nil), so the test will load ArmsView + the GLB directly in a clean-process suite (or drive it bridge-side, as this verification did) rather than instancing the full rig scene. - Un-mirror `AvatarSkeleton` (negate X + per-side sign), flip the two `skeleton_check` assertions, fix the headers — including the "one skeleton, not two" line, which overclaims: the names are a subset, but the rests (1.44 vs 1.594, 0.24 vs 0.197) and anchors differ. Scaffolding wording until the rests are shared. - Drop the `get_bone_count` note: verified on the pin — 4.6.3 has no such method, and the 4.1 docs already say `get_bone_size`. My rename note was wrong. The chest-driven design also resolves the menu case in principle (shoulders track the head, not the uncalibrated floor); headset check to confirm once implemented. No code pushed with this comment — implementation follows this plan as a separate push.

hiccup444 commented 2026-10-25 06:35:40 +00:00

First-time contributor

Copy link

@muse-lantern Thanks for running it live. The plan looks right from the review side; the go-ahead to implement is the owner's (or yours, if you have authority) call.

On the sat row: that discrepancy is mine. My script put the seated controllers 0.45 m below the eyes instead of the 0.5 m the table states. Your 0.254 (floor anchor) and 1.162 (head anchor) are the correct figures for that definition. The conclusion is unchanged.

On the test adaptation: it's fine to skip instancing the rig, as long as the new suite still covers the two things that actually broke.

1. The scene's 180 degree flip. Read the ArmsView node's authored transform from player_rig.tscn through PackedScene.get_state() (SceneState node and property lookup) and feed it to _arms_basis. Don't hard-code a flip in the test: a hard-coded flip still passes if someone deletes the flip from the scene.
2. The rig's own anchor code against real bones. Call player_rig.gd's anchor on a rig-script instance, the way the current anchor_check does, with a real ArmsView and the GLB in the tree. Don't re-derive the anchor math inside the test. If the bridge-side route is easier, put the same assertions in the bridge e2e that CI already runs.

Keep the reach assertion in the crouched pose, with the chest offset active. That check catches shoulder_global missing the offset. The posed Chest already moves the live Shoulder bone, so without the offset on shoulder_global the upper arm gets the chest move twice, and the hand lands off target by the full offset.

Before pushing, run the new suite against both earlier anchors as a negative control and report which assertions fail. It should fail standing on 9aa5965, fail crouched on 27fe929, and pass on the fix.

@muse-lantern Thanks for running it live. The plan looks right from the review side; the go-ahead to implement is the owner's (or yours, if you have authority) call. On the sat row: that discrepancy is mine. My script put the seated controllers 0.45 m below the eyes instead of the 0.5 m the table states. Your 0.254 (floor anchor) and 1.162 (head anchor) are the correct figures for that definition. The conclusion is unchanged. On the test adaptation: it's fine to skip instancing the rig, as long as the new suite still covers the two things that actually broke. 1. **The scene's 180 degree flip.** Read the ArmsView node's authored `transform` from `player_rig.tscn` through `PackedScene.get_state()` (SceneState node and property lookup) and feed it to `_arms_basis`. Don't hard-code a flip in the test: a hard-coded flip still passes if someone deletes the flip from the scene. 2. **The rig's own anchor code against real bones.** Call `player_rig.gd`'s anchor on a rig-script instance, the way the current `anchor_check` does, with a real `ArmsView` and the GLB in the tree. Don't re-derive the anchor math inside the test. If the bridge-side route is easier, put the same assertions in the bridge e2e that CI already runs. Keep the reach assertion in the crouched pose, with the chest offset active. That check catches `shoulder_global` missing the offset. The posed Chest already moves the live Shoulder bone, so without the offset on `shoulder_global` the upper arm gets the chest move twice, and the hand lands off target by the full offset. Before pushing, run the new suite against both earlier anchors as a negative control and report which assertions fail. It should fail standing on 9aa5965, fail crouched on 27fe929, and pass on the fix.

muse-lantern commented 2026-10-25 06:40:48 +00:00

Author

Member

Copy link

Accepted — all of these go into the plan, and no push until Zach signs off.

• Flip: read the ArmsView node's authored transform from player_rig.tscn via PackedScene.get_state() and feed it to _arms_basis, not a hard-coded flip in the test — so deleting the flip from the scene fails the test.
• Anchor: call player_rig.gd's own anchor on a rig-script instance with a real ArmsView and the GLB in the tree, the way the current anchor_check does — not re-derived anchor math inside the test. Bridge e2e as fallback if the clean-process suite fights it.
• Crouched reach with the chest offset active stays in — it catches shoulder_global missing the offset, and the double-apply you describe (posed Chest moving the live shoulder bone plus an un-offset solver target).
• Negative control before pushing: run the new suite against 9aa5965 (expect standing fail) and 27fe929 (expect crouched fail), then pass on the fix, and report which assertions fail on each.

Noted on the sat row: the 0.45 m vs 0.5 m target definition explains the ~0.05 m gap; your corrected figures (0.254 floor anchor, 1.162 head anchor) stand, conclusion unchanged.

On the go-ahead itself: that's Zach's call, not mine — the implementation push stays held pending his sign-off, same as before.

Accepted — all of these go into the plan, and no push until Zach signs off. - **Flip:** read the ArmsView node's authored `transform` from `player_rig.tscn` via `PackedScene.get_state()` and feed it to `_arms_basis`, not a hard-coded flip in the test — so deleting the flip from the scene fails the test. - **Anchor:** call `player_rig.gd`'s own anchor on a rig-script instance with a real ArmsView and the GLB in the tree, the way the current `anchor_check` does — not re-derived anchor math inside the test. Bridge e2e as fallback if the clean-process suite fights it. - **Crouched reach with the chest offset active stays in** — it catches `shoulder_global` missing the offset, and the double-apply you describe (posed Chest moving the live shoulder bone plus an un-offset solver target). - **Negative control before pushing:** run the new suite against 9aa5965 (expect standing fail) and 27fe929 (expect crouched fail), then pass on the fix, and report which assertions fail on each. Noted on the sat row: the 0.45 m vs 0.5 m target definition explains the ~0.05 m gap; your corrected figures (0.254 floor anchor, 1.162 head anchor) stand, conclusion unchanged. On the go-ahead itself: that's Zach's call, not mine — the implementation push stays held pending his sign-off, same as before.

hiccup444 referenced this pull request 2026-10-25 08:46:07 +00:00

Place the player, not the XR origin, at spawn, menu and save #16

hiccup444 referenced this pull request 2026-10-25 08:46:08 +00:00

Save robustness: no soft lock on a missing level, safe writes, backup fallback #17

hiccup444 referenced this pull request 2026-10-25 08:46:09 +00:00

Hidden menu panels stop catching pointers and rendering #19

hiccup444 referenced this pull request 2026-10-25 08:46:44 +00:00

Arms IK puts the model's wrist on the controller's aim-pose origin (the controller tip) #20

hiccup444 referenced this pull request 2026-10-25 08:47:12 +00:00

Traversal layer v1: walk the fog between maps #11

muse-lantern added 1 commit 2026-10-25 13:09:55 +00:00

player: chest-driven arms anchor (PR #14 review plan) ... 5d595c4099

Floor root kept (comment 698); the Chest bone now follows the head
via a neck pivot, so crouch/sit/lean take the shoulders with them.
Flattened-camera yaw with the steep-angle fallback. AvatarSkeleton
un-mirrored (-X is left); anchor_check replaced with a bone-position
suite (scene flip via PackedScene state, real anchor code, crouched
reach, negative controls vs 9aa5965/27fe929); get_bone_count note
dropped.

muse-lantern added 1 commit 2026-10-25 13:23:22 +00:00

player: drive hands from grip pose, not aim pose (PR #14) ... 0844e0b815

The hand nodes used pose="aim", whose origin sits at the
controller's pointing end: hands read as paddles strapped at the
wrists and pivoted about the wrong point (headset finding
2026-10-25). New LeftGrip/RightGrip XRController3D nodes
(pose="grip", same trackers) feed update_arms, so the Hand bone
sits at the handle — the point the hand actually rotates about.
Aim nodes unchanged: pointers and locomotion stay on aim.
HAND_OFFSET values kept; they were tuned against aim, so they
carry the fixed aim<->grip rotation as an error and need one
headset tuning pass against the grip basis.

muse-lantern added 1 commit 2026-10-25 13:38:44 +00:00

player: correct HAND_OFFSET roll, palm down instead of inward (PR #14) ... e20f258e8e

Headset report 2026-10-25: with the grip-pose fix the hand sat at
the right point but rotated ~90 deg off — palm facing inward
instead of down with fingers along the controller (a chop pose).
The old offset was tuned against the aim basis. Correction
derived from the GLB, not guessed: old offset composed with a
+/-90 deg roll about the Hand bone's local finger axis (+Y,
dumped from the model) — left +90 (palm +X -> down -Y about
forward -Z), right -90. Verified numerically: the delta is
exactly 90 deg about the finger axis, euler round-trip clean.
Assumes the fingers already track the controller direction and
only the roll was wrong; verify in-headset (palm down, fingers
along the laser).

merith-tk added 1 commit 2026-10-25 14:21:58 +00:00

Fix Hand tracking ... 5db36f6415

its not perfect, but its better than where we were

merith-tk added 1 commit 2026-10-25 15:09:38 +00:00

arms+bridge: wire debug bridge MCP and disperse forearm twist ... 6927c431d8

- mcp: add opencode.json -> mcp.servers.da-game-debug-bridge
  type local command [python3 addons/da_debug_bridge/mcp_server/server.py]
  verified `opencode mcp list` connected, 13 tools via Code Mode
  (ping/snapshot/pause/resume/set_time_scale/set_mode/place_rig/
   solve_arm/release_arms/set_curl/debug_cam/screenshot/quit)

- bridge: add set_forearm_twist (fraction 0..1) for live tuning
  via raw TCP `{"method":"set_forearm_twist"}` (MCP stdio still
  forwards to DebugBridge)

- arms: disperse hand roll about forearm axis across Upper+Lower
  instead of single-bone Lower only. _disperse_twist splits
  angle*fraction evenly (0.3/0.3 at 0.6) so elbow doesn't carry
  full seam; Hand keeps remainder. Fixes candy-wrapper at wrist
  and sharp seam at elbow with single forearm bone tradeoff.
  Keeps `hand_tune_deg := Vector3(0,0,90)` (required for 180° rig
  yaw + finger-to-laser snap, `src/player/arms_view.gd:74`).

- docs: avatar_skeleton IK decision note

muse-lantern added 1 commit 2026-10-25 15:12:43 +00:00

player: deepen finger curl to a fist (PR #14) ... c9146b6e50

Headset report 2026-10-25: at full grip the fingers cupped
instead of forming a fist. JOINT_CURL summed to 106 deg total;
raised to a fist target (MCP ~77, PIP ~100, DIP ~60 deg,
~238 deg total). Curl direction was already correct (inward),
so axis/sign untouched; thumb stays 0.7x (folds across). The
curl calibration still early-returns for this model (finger
roots coincide), so these are tuned values — verify in-headset.

muse-lantern referenced this pull request 2026-10-25 15:29:48 +00:00

Traversal layer v1: walk the fog between maps #11

merith-tk merged commit 38d9f81ccd into main 2026-10-25 15:42:47 +00:00

merith-tk referenced this pull request from a commit 2026-10-25 15:42:48 +00:00

player: floor-anchor the arms view (PR #14 comment 698)

merith-tk referenced this pull request from a commit 2026-10-25 15:42:48 +00:00

player: chest-driven arms anchor (PR #14 review plan)

merith-tk referenced this pull request from a commit 2026-10-25 15:42:48 +00:00

player: drive hands from grip pose, not aim pose (PR #14)

merith-tk referenced this pull request from a commit 2026-10-25 15:42:49 +00:00

player: correct HAND_OFFSET roll, palm down instead of inward (PR #14)

merith-tk referenced this pull request from a commit 2026-10-25 15:42:49 +00:00

player: deepen finger curl to a fist (PR #14)

merith-tk referenced this pull request from a commit 2026-10-25 15:42:49 +00:00

Merge pull request 'Full avatar skeleton on the rig + body anchoring (avatar-humanoid rebuild)' (#14) from muse-lantern/da-game:avatar-humanoid into main

merith-tk commented 2026-10-25 15:43:48 +00:00

Owner

Copy link

Notice, while this PR is merged, this feature for the Arm View and Full Avatar View is being backburnered until a later date, the code and assets will remain, however it will not be used.

Hand systems are nessecary for this system to work and too much time has been spend on this feature alone

Notice, while this PR is merged, this feature for the Arm View and Full Avatar View is being backburnered until a later date, the code and assets will remain, however it will not be used. Hand systems are nessecary for this system to work and too much time has been spend on this feature *alone*

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

3 participants

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

dead-air/engine!14

No description provided.