threepp for Python

threepp is a cross-platform C++ 3D library with the high-level API of three.js. This package exposes that high-level API to Python through a single pybind11 extension module, threepp.

You get the three.js scene graph — Scene, Mesh, geometries, materials, cameras, lights, OrbitControls, an animation loop — plus headless render-to-numpy for ML data generation, robotics cameras and machine vision.

import threepp as tp

scene = tp.Scene()
scene.add(tp.Mesh(tp.BoxGeometry(), tp.MeshStandardMaterial()))
scene.add(tp.AmbientLight())

camera = tp.PerspectiveCamera(75, 1.0, 0.1, 100)
camera.position.z = 5

canvas = tp.Canvas("demo")
renderer = tp.GLRenderer(canvas)
canvas.animate(lambda: renderer.render(scene, camera))

Install

Status: pre-1.0 / alpha. This is a young package and the API is still moving — it may change between releases. Releases are dated (a YYYY-MM-DD tag builds a YYYY.MM.DD wheel); pin the exact release you tested and check the release notes before upgrading.

Prebuilt wheels (Windows / Linux, CPython 3.10–3.14) are on PyPI — no build tools or system libraries needed:

pip install threepp

They ship the GL renderer, CPU PhysX physics (PhysxWorld, articulations, proprioceptive sensors — tp.HAS_PHYSX == True) and the Vulkan deferred renderer with its G-buffer AOVs (tp.HAS_VULKAN == True). Vulkan needs a Vulkan-capable GPU driver at runtime — on machines without one the wheel still imports and renders GL, and tp.vulkan_available() tells you which world you're in. macOS has no prebuilt wheel — pip install builds it from source there (GL-only; a C++ compiler and CMake are all it needs).

The scene editor is its own package — author physics-ready scenes, robots and sensors visually, then drive them from Python:

pip install threepp-editor      # or: pip install threepp[editor]  (Windows)
threepp-editor

(threepp[full] = [rl,editor]. The editor ships Windows wheels only; the extra is a no-op elsewhere, and pip install threepp-editor states the platform requirement honestly.)

Or build from source (needs a C++ compiler + CMake ≥ 3.21; pybind11 is fetched automatically):

pip install "git+https://github.com/markaren/threepp"

Physics in the wheel is CPU-only — the GPU runtime (PhysXGpu_64.dll) is 238 MB and stays out. Everything PhysxWorld does by default works. The GPU paths need the two GPU DLLs (PhysXGpu_64.dll + PhysXDevice64.dll) dropped next to the installed module (they are loaded lazily by name, never linked): that alone enables PhysxWorld(gpu_dynamics=True) and the direct-GPU RL batch (verified). Soft bodies (PhysxWorld.add_soft_body) run only on the GPU solver, so they need the same two DLLs.

Build

The module is built as part of threepp's CMake project, gated behind THREEPP_WITH_PYTHON. pybind11 is fetched automatically; the host Python is auto-detected (override with -DPython_EXECUTABLE=...).

cmake -S . -B build -DTHREEPP_WITH_PYTHON=ON \
      -DTHREEPP_BUILD_EXAMPLES=OFF -DTHREEPP_BUILD_TESTS=OFF
cmake --build build --target threepp_py

(On Windows, CMake defaults to the multi-config Visual Studio generator; the build works under it, but -G Ninja gives a faster single-config build.)

The built module (threepp.*.pyd / threepp.*.so) is placed in this python/ directory, so scripts here can import threepp. The examples add this directory to sys.path automatically, so they run from anywhere.

This is a GL-only build — it needs no Vulkan SDK, PhysX or CUDA, and works on any laptop. Off-screen rendering uses a hidden GLFW window (a real GL context), so a display is only required for the on-screen examples.

Examples

cd python
python examples/headless_render.py

The rest of examples/, by what it is about. One line per script, from its own module docstring.

Basics

Sensors

Physics and robots

Warp films — NVIDIA Warp simulation rendered live by threepp. Each is self-contained; what they share lives in examples/warp_common.py.

Granular soil and rovers: wheels carried by MLS-MPM grains (threepp.granular_mpm), checked against single-wheel terramechanics data.

Film kits — reusable pieces a film imports. Renderer-only (no Warp); each runs standalone too.

Inverse design — an optimiser, not an animator, decides the shape.

Probes — instruments, not demos: each answers one question and prints (or writes) the answer. They live in examples/probes/ with smoke_test.py above; see its README.

What's exposed

• Math (mutable value types): Vector2/3/4, Color, Euler, Quaternion, Matrix3/4, Box3. A hex int (0xff0000) or CSS/name string converts implicitly to a Color, so material.color = 0xff0000 just works.

• Scene graph: Object3D, Scene, Group, Mesh, InstancedMesh, Points, Line, LineSegments, Sprite. obj.position.x = 1 mutates in place, exactly like three.js; scene.add(a, b, c), traverse, get_object_by_name, children all work.

• Geometries: Box, Sphere, Plane, Cylinder, Cone, Capsule, Torus, TorusKnot, Circle, Ring, Icosahedron, Octahedron, plus the swept/extruded pair — TubeGeometry(curve, ...) and ExtrudeGeometry(shape, depth=...) / ShapeGeometry(shape).

• Curves and 2-D shapes: CatmullRomCurve3, LineCurve3 (both Curve3), LineCurve, SplineCurve, Path (both Curve2) and Shape — the sampling API (get_point, get_point_at, get_tangent, get_points, get_spaced_points, get_length) is shared by all of them. For long polylines use the array forms, which move no per-point Python objects: CatmullRomCurve3.from_array(points) / SplineCurve.from_array(points) take an (N, 3) / (N, 2) float array, and get_points_at(u), get_tangents_at(u), get_points_array(n), get_spaced_points_array(n) return (N, D) float32 arrays.

• Materials: MeshStandard, MeshPhysical, MeshPhong, MeshLambert, MeshBasic, MeshNormal, MeshDepth, Points, LineBasic, Sprite, Shadow — concrete fields, the shared base fields (opacity, transparent, side, …), and texture-map slots (map, normal_map, roughness_map, metalness_map, emissive_map, ao_map, alpha_map, …).

• Cameras: PerspectiveCamera, OrthographicCamera.

• Lights: Ambient, Directional, Point, Spot, Hemisphere, RectArea.

• Textures: Texture, TextureLoader (load(path, color_space=...)), with TextureWrapping / Filter / ColorSpace enums.

• Model loaders: ModelLoader (load(path) → Group, dispatches by extension: .obj / .gltf / .glb / .stl / .dae), plus OBJLoader, STLLoader, and GLTFLoader (load(path) → GLTFResult with .scene and .animations). All first-party — no Assimp/FBX/USD needed.

• Animation: AnimationMixer, AnimationClip, AnimationAction, the KeyframeTrack family (Vector/Quaternion/Number/Color), and the Loop / AnimationBlendMode / Interpolation enums — three.js' animation system. Play clips loaded from glTF, or build them procedurally:

  result = tp.GLTFLoader().load("model.glb")
  scene.add(result.scene)
  mixer = tp.AnimationMixer(result.scene)
  mixer.clip_action(result.animations[0]).set_loop(tp.Loop.REPEAT).play()
  
  clock = tp.Clock()
  canvas.animate(lambda: (mixer.update(clock.get_delta()),
                          renderer.render(scene, camera)))
  

• Rendering: Canvas (window / headless), GLRenderer (render, set_clear_color, read_pixels → numpy, save_frame, shadows, tone_mapping / tone_mapping_exposure), OrbitControls, Clock.

• HDR image-based lighting: RGBELoader().load("env.hdr") (or EXRLoader for .exr) → an equirect Texture; assign it to scene.environment (IBL on standard/physical materials) or scene.background. ToneMapping.ACESFilmic (etc.) keeps HDR highlights from clipping.

• Camera intrinsics / extrinsics: camera.projection_matrix, matrix_world_inverse, obj.matrix_world, get_world_quaternion, Matrix4.to_numpy() (→ (4, 4)), and Vector3.project(camera) / unproject for 3D↔2D — the building blocks for 6-DoF pose ground truth and 2D annotation.

• URDF robots: URDFLoader().load("robot.urdf") → a Robot (an Object3D) with set_joint_value(s), get_joint_range(s), num_dof, and forward kinematics (get_end_effector_transform() / compute_end_effector_transform).

• 2D text & SVG: FontLoader().default_font() (embedded — no font file), Text2D / Text3D (text meshes), TextSprite (billboard labels), and SVGLoader().parse(...) → a Group of filled meshes. Pair with an OrthographicCamera + auto_clear=False for a HUD.

• In-window UI: ImguiContext + the threepp.imgui submodule — Dear ImGui immediate-mode widgets (window/text/button/slider/checkbox/color/combo/…) for control panels. Works on both the GL and Vulkan renderers (tp.HAS_IMGUI).

• 3D Gaussian Splatting: SplatLoader.load_ply() reads the .ply files 3DGS optimisers emit (any SH degree, header-driven), SogLoader reads PlayCanvas SOG scans, and SplatCloud puts the result in the scene graph like any mesh — rendered by both the GL and Vulkan backends. is_splat_ply() discriminates splat files from mesh PLYs; submit_ranges exposes the chunk-LOD/culling mechanism (Vulkan).

• Mesh collision queries: BVH — BVH.intersects(a, b) (exact triangle-triangle, early-exit), BVH.intersect(a, b) → an (N, 3) array of points on both surfaces, BVH.distance(a, b, max_distance=...), and raycast. Build from a BufferGeometry or from numpy (build_arrays(positions, indices)). Every query releases the GIL; intersects_many, intersect_many and distance_many evaluate a candidate x target grid in parallel in one call:

  volumes = [tp.BVH() for _ in tracks]
  for bvh, track in zip(volumes, tracks):
      g = tp.TubeGeometry(tp.LineCurve3(*track), radius=5.0, radial_segments=8)
      bvh.build_arrays(g.get_attribute("position"), g.get_index())
  
  hits = tp.BVH.intersects_many(candidates, volumes)      # (C, T) bool
  when = tp.BVH.intersect(candidates[0], volumes[2])[:, 1]  # y of each contact point
  

• Layers: obj.layers / camera.layers (set, enable, disable, test) — the three.js visibility mask.

• Rigid-body physics (when built with PhysX, see below): PhysxWorld + RigidBody — add Meshes as dynamic/static bodies (box/sphere/capsule, convex hull, or triangle mesh), step(dt), and the bound meshes follow the simulation (tp.HAS_PHYSX). Reduced-coordinate Articulations (URDF robots with joints, limits and drives), PhysxVehicle and GPU SoftBodys are bound too.

• Proprioceptive sensors (PhysX builds): Imu (gyroscope + accelerometer), JointEncoder, ContactSensor and ForceTorqueSensor — attached to scene nodes, driven from the physics step loop, with configurable seeded NoiseModels. See Proprioceptive sensors below.

• Vulkan deferred renderer + G-buffer AOVs (when built with Vulkan, see below): VulkanRenderer.render_aov(scene, camera, aov) returns a deferred G-buffer attachment as (H, W, 3) uint8 — 'rgb', 'normals', 'segmentation' (per-instance ids), 'albedo', 'motion' — and the lossless typed family reads the native attachments: read_depth (metric (H, W) float32), read_instance_ids (raw uint32 ids), read_normals_float, read_motion, read_aovs_typed. This is the "labels for free" path for synthetic-data generation.

• FFT ocean (renders under the Vulkan backend): tp.Ocean(size=1000.0) is a ready-made 3-cascade Phillips/FFT-displaced water surface — waves, foam, and transmission. Add it to a scene and render. Tune ocean.params (wind / cascades / choppiness), pack vertex density toward any focus with warp_toward(x, z), add add_foam_disturbance(...) splats, and read the wave height on the CPU with sample_height(x, z) (lazy opt-in: the first call enables the GPU->CPU mirror and may return 0 — values are live from the next rendered frame). The low-level tp.DisplacedMesh (own geometry + material) is exposed too. Not tied to any "hero" object — see examples/vulkan_ocean.py.

• Pure-Python helpers that ship in the package beside the native module: threepp.cuda_interop (a mesh's vertex buffers as CUDA arrays an external producer writes in place) and threepp.torch_frames (the Vulkan renderer's images as live torch tensors), both NVIDIA-only; threepp.terrain_deform (Sumner/O'Brien/Hodgins deformable ground with a Bekker-Wong bearing, batched in Warp) and threepp.granular_mpm (MLS-MPM granular soil with two-way rigid wheel and plate colliders, in Warp); threepp.urdf (a URDF straight into a PhysX articulation) and threepp.rl. The Warp modules import Warp only on first use, so import threepp never needs it.

Naming follows Python conventions (snake_case methods/properties), e.g. camera.update_projection_matrix(), renderer.set_clear_color(...).

Tests

A pytest suite under tests/ covers the whole surface — math, scene graph, geometries, materials, cameras, lights, textures, loaders and headless rendering — including regression tests that pin the pybind11 virtual-base workaround (they crash the interpreter if it ever regresses).

pip install pytest numpy pillow
cd python
pytest

examples/probes/smoke_test.py is a dependency-light standalone alternative that prints ALL OK.

Type stubs (IDE autocomplete)

Stubs ship alongside the module as a PEP 561 stub package, mirroring the native module's own structure:

threepp/py.typed                 # PEP 561 marker (makes the types count once installed)
threepp/threepp/__init__.pyi     # the threepp.threepp extension module
threepp/threepp/imgui.pyi        # the threepp.threepp.imgui submodule
threepp/threepp/editor.pyi       # the threepp.threepp.editor submodule — HAND-MAINTAINED

This replaces the old flat threepp/threepp.pyi, which could not describe the imgui submodule. The threepp/threepp/ directory sits next to threepp.<abi>.pyd / .so but does not shadow it at import time: it has no __init__.py, so it is only a namespace-package candidate, and Python's import machinery prefers the extension-module loader within the same directory.

Regenerate after changing the bindings:

pip install -r python/requirements-stubs.txt
cmake --build build --target threepp_stubs

Or, once the module is built into the package dir, standalone:

python python/scripts/gen_stubs.py

pybind11-stubgen is pinned in requirements-stubs.txt because its output layout and formatting change between releases — an unpinned upgrade is what silently reshaped the stubs from a flat file into this package. Bump the pin deliberately, regenerate, and review the diff.

scripts/gen_stubs.py wraps the generator so the result is reproducible: it fixes the flags, repairs any keyword-named binding (see below), fails if the emitted stub does not parse, and fails if regeneration dropped symbols the committed stubs declare. The bare pybind11-stubgen CLI exits 0 even when it writes a stub no type checker can read, which is how the previous stubs went stale unnoticed.

Regenerate from a full-featured build

The stubs describe whatever the built module exposes, so regenerate from a build with the same features as the last one — Vulkan, PhysX, FSR. A default or GL-only rebuild drops VulkanRenderer, the PhysX world and every sensor binding, and the result still parses, so nothing else would catch it.

gen_stubs.py guards this: it snapshots the committed stubs' symbols before regenerating and fails if any disappeared, naming them. If a reduction is genuinely intended, say so:

python python/scripts/gen_stubs.py --allow-removals

editor.pyi is hand-maintained

Every other stub here is generated; editor.pyi is written by hand, and gen_stubs.py restores it verbatim after each run (HAND_MAINTAINED).

threepp.editor is served by two modules built from the same binding sources: this wheel, and the editor app's embedded interpreter. Only the editor compiles src/bind_editor_physics.cpp — RigidBody, SoftBody, Articulation and the three *_from_object lookups are handles onto a live PhysicsPlaySession, which nothing outside a running editor has, so the wheel does not bind them. The stub has to describe the union for a script author's completion to be right, and a stub generated from the wheel alone would delete the physics half (41 symbols).

So: change bind_editor_physics.cpp, and edit editor.pyi to match. The restore happens after the produced-files check, so a threepp.editor that disappeared from the module still fails the run rather than being papered over by the committed copy.

The union stays in the source tree for the editor's Pylance integration, but it does not ship in the wheel: the wheel carries py.typed, so an installed stub is certified to a type checker, and certifying 40-odd names that raise AttributeError on a pip install would be lying. The wheel instead installs wheel-stubs/editor.pyi — just SplinePath and spline_from_object, the two names bind_editor.cpp actually provides there. A name added to bind_editor.cpp belongs in both stubs; a name added to the editor-only TUs belongs only in the union.

Keyword-named bindings

A bound name that is a Python keyword cannot be written in Python source or in a stub, and one is enough to make the whole stub unparseable. Two such names have been renamed at the binding site:

No binding needs repair today. gen_stubs.py keeps a general repair pass as a standing guard, since pybind11 will emit the same breakage from any future C++ identifier that collides with a Python keyword.

A few signatures show ... instead of a real parameter type (e.g. Vector3.apply_matrix4). That is pybind11 baking a signature before the argument's C++ type was registered; the fix is binding declaration order, not the stub generator.

Vulkan deferred AOVs (synthetic data)

The deferred (RasterFirst) Vulkan renderer writes a full G-buffer every frame — world normals, optical flow, per-instance segmentation ids, albedo, depth. The binding exposes those attachments as numpy, so a scene authored in Python yields ground-truth labels directly from the render, with no manual annotation:

import threepp as tp
assert tp.HAS_VULKAN

canvas = tp.Canvas("aov", width=512, height=384, headless=True, vsync=False)
renderer = tp.VulkanRenderer(canvas)            # deferred / RasterFirst

# ... build scene + camera ...
out = renderer.render_aovs(scene, camera, ["rgb", "normals", "segmentation"])
rgb, normals, seg = out["rgb"], out["normals"], out["segmentation"]   # (H, W, 3) uint8
depth = renderer.read_depth(scene, camera)                           # (H, W) float32, metres

Enable the Vulkan backend at configure time (needs the Vulkan SDK + a Vulkan-capable GPU). The GL renderer stays available in the same module:

cmake -S . -B build -DTHREEPP_WITH_PYTHON=ON -DTHREEPP_WITH_VULKAN=ON
cmake --build build --target threepp_py

tp.HAS_VULKAN reports whether the backend was compiled in.

Two readback paths. The 8-bit visualisation AOVs (render_aov / render_aovs) come out as (H, W, 3) uint8 via the renderer's debug-resolve pass: normals as n*0.5+0.5, segmentation as per-id hashed colours, albedo as linear albedo (so it looks gamma-/hue-off as a viewed image, but the data is the real linear base colour). The lossless path copies the native G-buffer attachments straight to host memory: read_depth → metric (H, W) float32 (full-precision native D32 read, verified accurate against known distances), read_instance_ids → raw (H, W) uint32 per-instance ids, read_normals_float, read_motion, and read_aovs_typed([...]) for several in one render. Driving is via the deferred frame-model under the hood (submit/present is deferred to the canvas frame-end callback, so each render* repeats a few frames to make the MAILBOX readback deterministic — tune with set_flush_frames).

Headless / display-less machines (cloud GPUs). A headless=True canvas needs no display at all: the Vulkan renderer creates its surface via VK_EXT_headless_surface (supported by NVIDIA's Linux driver and Mesa) instead of a window surface, and on Linux with no DISPLAY/WAYLAND_DISPLAY the canvas skips the window system entirely (GLFW Null platform). That is exactly the Colab/EC2-style setup — compute-only NVIDIA kernel modules, no X server the GPU can present to — where the wheel's synthetic-data path (render_aov / read_depth / read_rgb_pixels) runs unmodified; presenting to a headless surface is simply a no-op. Where the ICD lacks the extension (e.g. NVIDIA's Windows driver), a headless canvas falls back to the previous behaviour — a hidden window with a real surface — and logs that it did so. Set THREEPP_GLFW_PLATFORM=null to force the window-system-free path on any OS (useful for reproducing the cloud setup locally).

In-window UI (Dear ImGui)

threepp integrates Dear ImGui; the binding exposes it as ImguiContext plus the threepp.imgui submodule of immediate-mode widgets. Issue widgets each frame inside a draw callback, and call ui.render(draw) in the animate loop after renderer.render(...):

ui = tp.ImguiContext(canvas)        # create AFTER the GLRenderer

def draw():
    tp.imgui.begin("Controls")
    changed, rough = tp.imgui.slider_float("roughness", material.roughness, 0.0, 1.0)
    if changed: material.roughness = rough
    _, material.wireframe = tp.imgui.checkbox("wireframe", material.wireframe)
    if tp.imgui.button("reset"): camera.position.set(0, 2, 6)
    tp.imgui.end()

def animate():
    controls.enabled = not ui.want_capture_mouse   # don't orbit while over the panel
    controls.update()
    renderer.render(scene, camera)
    ui.render(draw)

canvas.animate(animate)

Value-returning widgets give back (changed, new_value) tuples; button returns True on the clicked frame. tp.imgui.show_demo_window() opens the full ImGui gallery. tp.HAS_IMGUI reports availability.

It works on both renderers — pass the renderer so the backend matches: ImguiContext(canvas, gl_renderer) (or ImguiContext(canvas)) for GL, and ImguiContext(canvas, vulkan_renderer) for the Vulkan overlay (recorded into the deferred frame after the scene). Create the ImguiContext after the renderer. See examples/vulkan_ui.py for the Vulkan version.

One caveat: only one ImguiContext should be alive at a time (Dear ImGui has a single global context) — don't keep a GL and a Vulkan one simultaneously.

Rigid-body physics (PhysX)

threepp ships a scene-graph-integrated PhysX wrapper; the binding exposes it as PhysxWorld (+ a RigidBody handle). Add meshes as bodies, step(dt) each frame, and the bound meshes' transforms follow the simulation. It's pure CPU — no canvas or renderer required, so it works headless:

import threepp as tp
assert tp.HAS_PHYSX

world = tp.PhysxWorld(gravity=tp.Vector3(0, -9.81, 0))

floor = tp.Mesh(tp.BoxGeometry(20, 1, 20), tp.MeshStandardMaterial())
floor.position.y = -0.5
world.add_static(floor)                 # static collider

box = tp.Mesh(tp.BoxGeometry(1, 1, 1), tp.MeshStandardMaterial())
box.position.set(0, 5, 0)
body = world.add(box, density=200)      # dynamic body, auto-bound to the mesh
body.add_impulse(tp.Vector3(2, 0, 0))

for _ in range(120):
    world.step(1 / 60)                   # box.position now follows the sim

Shapes are inferred from Box/Sphere/Capsule geometry; add_dynamic_convex (convex hull), add_static_trimesh / add_static_trimesh_tree (exact triangles, e.g. an imported glTF environment), and add_instanced (one body per InstancedMesh instance) cover the rest. RigidBody exposes pose, linear/angular velocity, add_force / add_impulse, damping, and a kinematic mode (set_kinematic + set_kinematic_target). on_pre_substep / on_post_substep hooks fire around each fixed substep.

Enable PhysX at configure time. It comes from vcpkg (unofficial-omniverse-physx-sdk), so configure with the vcpkg toolchain:

cmake -S . -B build -DTHREEPP_WITH_PYTHON=ON \
      -DCMAKE_TOOLCHAIN_FILE=$VCPKG_ROOT/scripts/buildsystems/vcpkg.cmake
cmake --build build --target threepp_py

(Or, to reuse an existing install without the toolchain, pass -Dunofficial-omniverse-physx-sdk_DIR=<vcpkg_installed>/x64-windows/share/unofficial-omniverse-physx-sdk; the build then stages the PhysX runtime DLLs next to the module.) tp.HAS_PHYSX reports whether it was compiled in. Reduced-coordinate articulations (robots) are bound too — Articulation / ArticulationLink, built from a URDF in one call. So are PhysxVehicle, a drivable four-wheel PxVehicle2 whose road can be replaced per wheel by your own ground model, and SoftBody (add_soft_body, which needs gpu_dynamics=True and so a CUDA GPU).

Combined with the Vulkan AOVs, this is the dynamic half of the synthetic-data story: physics gives you moving scenes, the G-buffer gives you per-frame segmentation / depth / optical-flow labels for free.

Proprioceptive sensors (IMU)

On top of the PhysX world, threepp exposes a proprioceptive sensor suite for robotics: Imu (gyroscope + accelerometer), JointEncoder, ContactSensor and ForceTorqueSensor, each with seeded noise models. The walkthrough below uses the Imu. A sensor rides the scene graph: you attach it to an Object3D, and that node's world frame is the measurement frame. Register it with the world and it is sampled from the physics step loop — one clean sample per fixed substep, timestamped with the accumulated sim time — so the sampling fidelity is right for lock-step co-simulation (e.g. a future ArduPilot SITL).

import threepp as tp
assert tp.HAS_PHYSX

world = tp.PhysxWorld(gravity=tp.Vector3(0, -9.81, 0), fixed_timestep=1/240)
world.add_static(floor_mesh)                 # top face at y=0
body = world.add(box_mesh, density=200)      # dynamic body

mount = tp.Group()                           # IMU node, offset from the CoM
mount.position.set(0.3, 0.2, 0.1)
box_mesh.add(mount)

imu = tp.Imu(mount)                          # rate_hz=0 -> sample every substep
world.register_sensor(imu)                   # AFTER adding the body it rides

for _ in range(720):
    world.step(1/240)

for s in imu.drain():                        # oldest-first; empties the buffer
    print(s.t, s.angular_velocity, s.linear_acceleration)
data = imu.drain_array()                     # or a (N, 7) numpy [t, gx,gy,gz, ax,ay,az]
last = imu.latest()                          # most recent sample (survives drain)

Contract. Attach to a node; set rate_hz (0 = every physics substep, else the sensor sub-samples). Read non-blocking: latest() is the most recent sample, drain(out) / drain_array() move everything accumulated since the last drain out of a bounded ring buffer (default 2048; oldest dropped on overflow). register_sensor resolves the rigid body by walking up from the attachment node to the nearest ancestor added to the world — it raises immediately if there is none.

IMU units & frames. Both readings are expressed in the sensor (node) frame:

Specific force is what a real accelerometer reads (proper acceleration minus gravity, threepp is Y-up so g = (0, −9.81, 0)), so a level body at rest reads (0, +9.81, 0) and a body in free fall reads ~0. When the node is offset from the centre of mass, the lever-arm terms α × r + ω × (ω × r) are included (a_com and α come from finite-differencing the PhysX velocities across samples; the first sample after attach/reset() emits zero to avoid a start-up spike).

Noise. imu.gyro_noise and imu.accel_noise are NoiseModels with per-axis continuous-time densities — white_noise_density [X/√Hz], random_walk [X/(s·√Hz)] bias instability, and a constant constant_bias [X] — plus a seed (deterministic given seed + call sequence). Defaults are consumer MEMS-class; set every field to zero for a perfect sensor. Change the noise, then call imu.reset() (which also re-arms the finite-difference after an episode reset).

See examples/imu_demo.py for the full headless demo.

Notes for maintainers

threepp uses virtual inheritance for some classes (Mesh/Points/Line derive from Object3D virtually; concrete materials derive from Material virtually). pybind11 mishandles pointer adjustment across a virtual base, which corrupts memory. The bindings work around this by:

• never letting pybind up-cast a derived Python object to a shared_ptr of a virtual base — as_object3d / as_material cast to the concrete type and let the C++ compiler do the up-cast;
• binding the inherited Object3D / Material members on each concrete leaf with concrete member pointers (&T::field) and T& lambdas, instead of inheriting the (broken) base bindings.

See src/bind_objects.cpp (bind_object3d_api) and src/bind_materials.cpp (bind_material_base_fields).