Skip to content

DisplayXR v2.36.0

Choose a tag to compare

@github-actions github-actions released this 11 Nov 18:37
· 33 commits to main since this release

[2.36.0] - 2026-11-11

Fixed

  • Hands: under a camera-centric rig a hand at the glass showed behind zero parallax and about 3x too small. The camera rig anchored the glass at the nominal viewer distance (about 0.6 m) in front of the camera at metric scale. In a scene converging at 0.5 m with a ~0.58 m virtual screen, that put a hand in front of the glass 6–10 cm behind the screen plane. The glass now sits on the camera's convergence plane (1 / invConvergenceDistance in front of it, zero parallax), and joints are scaled by the camera's virtual screen height there over the canvas height. A display rig has always behaved this way, and the runtime now does the same for chained camera rigs (XR_DXR_view_rig spec 5). The C# fallback DisplayXRHandRigMath.CameraRig gains an overload taking invConvergenceDistance and tan(vFOV/2). The 3-argument form is the infinite-convergence mapping, which is unchanged. Unit-tested in Tests~/HandRigMath against the runtime rule. On a spec-4 runtime (v2.47.1–v2.59.x), chained camera-rig joints keep the old placement until the runtime is updated.

  • Hands: a camera-centric rig put every joint behind its camera; joints ignored an off-centre canvas (#439). On DisplayXRCamera (the Basic Scene sample) a hand 10 cm in front of the glass came back at z = -0.298 relative to the rig camera, so the Ghost Hand was never visible though tracking and pinch worked. The camera rig's camera is the nominal viewer, not the display plane, and the plug-in scaled the joints by the convergence-plane framing with no viewer offset. Joints now map exactly as an eye at that point would — the runtime's rule for rig-scoped hand joints (XR_DXR_view_rig spec 4): minus the canvas centre, then display rig × perspectiveFactor · virtualDisplayHeight / canvas height, camera rig z − nominalViewerZ unscaled (the panel, and a hand in front of it, sit about half a metre in front of the camera). On a runtime advertising XR_DXR_view_rig ≥ 4 (v2.47.1+) the provider chains the view locate's rig (identity pose) on xrLocateHandJointsEXT and the runtime returns rig-frame joints — the only path that is right for a window off the panel centre, since runtime v2.56 places an unchained display-plane hand source at the rig's image of the display centre. Older runtimes get the same table in C# (DisplayXRHandRigMath, unit-tested in CI against the shared core's eye mapping). New: DisplayXRHandTracking.ResolveMapping / CurrentMapping / JointsInRigFrame, native export DisplayXR_HandTracking_IsRigFrame; ResolvePoseScale / PoseScaleOverride keep their meaning.

  • Ghost Hand sample: the model was about 2x too big for its joints (#440). A hand source that maps physical motion with a gain below 1 (a relative mapping into the display volume, about 0.48 m/m by default) makes the joints about half life-size, while the sample drove only the rotations of a real-size model. DisplayXRGhostHand now sizes the model from the tracked hand (wrist → middle-proximal against the model's own wrist → middle-knuckle bone), smoothed over sizeSmoothingSeconds; matchTrackedHandSize = false restores the old real-size behaviour. The joint-sphere fallback is unchanged.

  • main's committed Windows DLL predated hand tracking (#441). It had no DisplayXR_HandTracking_* exports, so a from-main Windows build never started hand tracking (releases were fine: CI builds the shipped DLL). Refreshed from this change's CI build.

  • package.json now declares com.unity.ugui (#442). DisplayXRLocal2D and DisplayXRTextIme use it; a bare project without uGUI failed to compile the package. 1.0.0 is Unity 2022.3's version; Unity 6's 2.0.0 satisfies it.

  • Native log: the session-created hand-tracking line no longer reads app requested=0 (#445). It now says whether trackers are created now or on XR Hands Start, logs which frame the joints come back in (rig frame / display-plane metres) once per session, and the first-seen wrist line names its units.

  • Hands: a camera-centric scene still placed the joints behind its camera (#446 follow-up, #439). On the partner's rerun of a DisplayXRCamera scene the wrist came back at z = -0.146 relative to the camera, and the runtime logged rig-scoped hand joints (display rig). The hand locate rebuilt its rig descriptor from the provider's live tunables, which are zeroed at every session start and not pushed by the C# driver while the boot splash is up — so for those seconds both the views and the hands used the provider's default display rig, and display-frame joints read as camera-relative sit behind the camera. The hand locate now chains a snapshot of exactly the rig the latest primary view locate used; a new export DisplayXR_HandTracking_RigKind (DisplayXRHandTracking.JointRig) says which rig that was, and the subsystem reports the hand untracked while it does not match the scene's rig (DisplayXRHandRigMath.JointRigFitsScene, unit-tested). The native joints in the RIG frame (... chained ...) line now prints on every change of the chained rig as well as once per session: it keyed on rig frame yes/no, so the default → app rig switch a few seconds into Play went unlogged and the only line said display rig.