Pre-release

This release is a pre-release and may not be stable for production use.

Swift

GitHub repository | Documentation (not yet published)

Swift is a light-weight browser-based animation visualizer which provides:

visualisation of mesh objects (Collada, STL, OBJ, glTF/GLB, PLY, VRML/WRL, and PCD files) and primitive shapes;
visualisation of multi-link robots created with the Robotics Toolbox for Python;
interactive UI controls (sliders, buttons, and more) for driving a scene from the browser;
recording and saving a video of the simulation;
source code which can be read for learning and teaching;

Built using Python and Javascript, Swift is cross-platform (Linux, MacOS, and Windows) while also leveraging the ubiquity and support of these languages.

A Panda arm following a slider-controlled target box in Swift

Swift provides robotics-specific functionality for rapid prototyping of algorithms, research, and education. Through the Robotics Toolbox for Python, Swift can visualise over 30 supplied robot models: well-known contemporary robots from Franka-Emika, Kinova, Universal Robotics, Rethink as well as classical robots such as the Puma 560 and the Stanford arm. Swift is under development and will support mobile robots in the future.

What's new in 2.0

Swift's browser frontend has been rebuilt from scratch as modern, dependency-free ES modules (no bundler, no framework, current three.js) — see this release's changelog for the full list. For existing users:

a bottom-left playback panel with a pause/play button and a realtime-speed selector (Max/1x/0.5x/0.25x) — see Playback controls below;
WebRTC support has been removed (comms="rtc", the vision install extra) — it had no live-camera use case and wasn't providing anything a plain WebSocket doesn't already handle for the normal desktop/browser setup this simulator targets;
env.add() is now four explicit methods — add_shape(), add_ui(), add_assembly(), add_robot() — one entry point per kind of thing, no type-checking required. env.add() still works and dispatches to these, kept for backward compatibility;
add_robot()/add_assembly() return an AssemblyHandle that owns that instance's live joint state (handle.q, handle.qd) — the robot model (or bare forward-kinematics function, for add_assembly()) stays plain and shareable, driven functionally (panda.fkine(handle.q), panda.jacobe(handle.q)). Setting robot.q/robot.qd directly still works but is deprecated;
any shape, assembly, or robot can take a per-step callback — callback=lambda t, values: ... — invoked by env.step() with the current sim time and any named UI element values, returning the new pose (shape) or q (assembly/robot). Removes the need to hand-write a loop body that mutates pose/q each step;
UI elements take an optional name=, and their current value is kept in env.values — available to any callback as values[name] without writing a per-element setter function;
env.show() prints the current display list (every shape/assembly/robot/UI element, with its id and name if given) for debugging.

Examples

These build up from the simplest possible scene to a fully interactive one. All are in examples/ and runnable as-is.

Render a box

The simplest possible Swift scene: one shape, no motion.

import spatialgeometry as sg
from spatialmath import SE3
from swift import Swift

env = Swift()
env.launch(realtime=True)

box = sg.Cuboid([0.2, 0.2, 0.2], pose=SE3(0, 0, 0.1), color=[0.2, 0.4, 1.0, 1.0])
env.add_shape(box)

env.hold()  # keep the browser tab open

Render a box, with sliders to move it

Named sliders (name=...) push their current value into env.values; a per-step callback reads it from there and returns the box's new pose — no per-slider setter function, no manual pose assignment in the loop. The Z slider controls height above the floor (the box's bottom face, not its centre), so it's never possible to clip the box through the ground:

import time
import spatialgeometry as sg
from spatialmath import SE3
from swift import Swift, Slider

env = Swift()
env.launch(realtime=True)

SIDE = 0.2
box = sg.Cuboid([SIDE, SIDE, SIDE], pose=SE3(0, 0, SIDE / 2), color=[0.2, 0.4, 1.0, 1.0])


def box_pose(t, values):
    # z is height above the floor, not the box centre
    return SE3(values["x"], values["y"], values["z"] + SIDE / 2)


env.add_shape(box, callback=box_pose)

env.add_ui(Slider(lambda v: None, min=-0.5, max=0.5, step=0.01, value=0.0, desc="Box X", unit="m"), name="x")
env.add_ui(Slider(lambda v: None, min=-0.5, max=0.5, step=0.01, value=0.0, desc="Box Y", unit="m"), name="y")
env.add_ui(Slider(lambda v: None, min=0.0, max=0.6, step=0.01, value=0.0, desc="Box Z", unit="m"), name="z")

while True:
    env.step(0.05)
    time.sleep(0.05)

A 2-link arm, built from raw shapes (no robot model)

Before reaching for a full robot model, it's worth seeing what one actually automates. add_assembly(fk, parts) takes a pure forward-kinematics function — given the assembly's current q, return one world pose per part — plus the parts themselves, and swift builds and owns the handle for you. Here two elongated cuboids stand in for the links of a 2-revolute-joint arm:

import time
import numpy as np
import spatialgeometry as sg
from spatialmath import SE3
from swift import Swift, Slider

env = Swift()
env.launch(realtime=True)


class TwoLinkArm:
    """A pure kinematic model: two links, two revolute joints about z."""

    def __init__(self, L1=0.3, L2=0.25, thickness=0.03):
        self.L1 = L1
        self.L2 = L2
        self.link1 = sg.Cuboid([L1, thickness, thickness], color=[0.8, 0.2, 0.2, 1.0])
        self.link2 = sg.Cuboid([L2, thickness, thickness], color=[0.2, 0.4, 1.0, 1.0])

    def part_poses(self, q) -> list[SE3]:
        """World pose of each link, purely as a function of q. Each
        cuboid's local origin sits at its own proximal (joint) end, so
        Tx(length / 2) places its centre correctly."""
        joint1 = SE3.Rz(q[0])
        joint2 = joint1 * SE3.Tx(self.L1) * SE3.Rz(q[1])
        return [joint1 * SE3.Tx(self.L1 / 2), joint2 * SE3.Tx(self.L2 / 2)]


arm = TwoLinkArm()
handle = env.add_assembly(
    arm.part_poses,
    [arm.link1, arm.link2],
    q0=[0.0, 0.0],
    callback=lambda t, values: [values["q1"], values["q2"]],
)

env.add_ui(Slider(lambda v: None, min=-np.pi, max=np.pi, step=0.01, value=0.0, desc="Joint 1", unit="rad"), name="q1")
env.add_ui(Slider(lambda v: None, min=-np.pi, max=np.pi, step=0.01, value=0.0, desc="Joint 2", unit="rad"), name="q2")

while True:
    env.step(0.05)
    time.sleep(0.05)

TwoLinkArm is the kinematic model — shareable, and it knows nothing about its own current configuration. env.add_assembly() builds the piece that does: an AssemblyHandle owning live q, exactly what add_robot() (below) returns for a real rtb.Robot — same class either way. The leap from here to a full robot model is then just: a kinematic model with more than two links, described by roboticstoolbox instead of by hand.

Render a Panda

import roboticstoolbox as rtb
from swift import Swift

env = Swift()
env.launch(realtime=True)

panda = rtb.models.Panda()
handle = env.add_robot(panda)
handle.q = panda.qr

env.hold()  # keep the browser tab open

A box driven by pure kinematics (no interaction)

Not every scene needs sliders — a pose can just as easily be a function of time. Here a box orbits the origin on a circle of radius 3*W (W = box width), while the plane of that circle slowly tilts about the x-axis. Swift owns t and calls the callback each step, so there's no manual pose assignment in the loop:

import time
import spatialgeometry as sg
import spatialmath as sm
from swift import Swift

env = Swift()
env.launch(realtime=True)

W = 0.1
box = sg.Cuboid([W, W, W], color=[0.2, 0.4, 1.0, 1.0])


def orbit(t, values):
    return sm.SE3.Rx(t / 10) * sm.SE3.Rz(t) * sm.SE3.Tx(3 * W)


env.add_shape(box, callback=orbit)

dt = 0.02
while True:
    env.step(dt)
    time.sleep(dt)

Tx(3*W) places the box at the orbit radius; Rz(t) spins it around the circle; Rx(t/10), applied last (so in the world frame, after the orbit is computed), tilts the whole orbital plane about x — ten times slower than the orbit itself.

Worth knowing: the ground plane is solid and opaque, so once the tilt carries the box below z=0 it's genuinely hidden underneath the ground from a normal above-ground viewpoint — the same as burying a real box and looking down at the dirt, not a rendering glitch. This particular orbit has no z-offset, so it dips underground for part of every cycle by construction; add a z-offset to Tx/wrap it in a translation if you'd rather the whole orbit stayed above the floor.

Panda arm follows a target, positioned by sliders

The most complete example: a target box is positioned by three named sliders, and on every step the arm runs resolved-rate motion control (rtb.p_servo) towards wherever that box currently is — move a slider, the box moves, and the arm continuously chases it. Both the box's pose and the arm's q are computed by per-step callbacks reading from env.values, so the whole thing runs off a plain env.step() loop with no pose/q mutation inside it:

import time
import numpy as np
import roboticstoolbox as rtb
import spatialgeometry as sg
import spatialmath as sm
from swift import Swift, Slider

env = Swift()
env.launch(realtime=True)
dt = 0.05

panda = rtb.models.Panda()
handle = env.add_robot(panda)
handle.q = panda.qr

SIDE = 0.05
X0, Y0, Z0 = 0.5, 0.0, 0.3  # initial target position; z is height above the floor

# The gripper points straight down at the target
target_orientation = sm.SE3.Rx(np.pi)

target = sg.Cuboid(
    [SIDE, SIDE, SIDE],
    pose=sm.SE3(X0, Y0, Z0 + SIDE / 2) * target_orientation,
    color=[0.2, 0.4, 1.0, 1.0],
)


def target_pose(t, values):
    return sm.SE3(values["x"], values["y"], values["z"] + SIDE / 2) * target_orientation


env.add_shape(target, callback=target_pose)


def track_target(t, values):
    v, _ = rtb.p_servo(panda.fkine(handle.q), target.T, gain=1.0, threshold=0.01)
    qd = np.linalg.pinv(panda.jacobe(handle.q)) @ v
    return handle.q + qd * dt


handle.callback = track_target

env.add_ui(Slider(lambda v: None, min=0.2, max=0.7, step=0.01, value=X0, desc="Target X", unit="m"), name="x")
env.add_ui(Slider(lambda v: None, min=-0.4, max=0.4, step=0.01, value=Y0, desc="Target Y", unit="m"), name="y")
env.add_ui(Slider(lambda v: None, min=0.05, max=0.6, step=0.01, value=Z0, desc="Target Z", unit="m"), name="z")

while True:
    env.step(dt)
    time.sleep(dt)

Embed within a Jupyter Notebook

To embed within a Jupyter Notebook cell, use the browser="notebook" option when launching the simulator — any of the examples above work the same way, just swap the env.launch(...) line:

env.launch(realtime=True, browser="notebook")

Playback controls

Every non-headless session shows a small panel bottom-left of the browser view:

a pause/play button (||/▶) — also bound to the spacebar;
a realtime-speed selector (Max/1x/0.5x/0.25x) — 1x matches env.launch(realtime=True), Max matches the default (realtime=False, uncapped). env.launch(realtime=0.5) (a specific float, not just True/False) sets an initial speed directly.

Recording video

Any scene can be recorded straight from Python — call env.start_recording(...) around the part you want captured, env.stop_recording() when done. The .webm file downloads automatically once encoding finishes:

env.start_recording("my_recording", framerate=20, format="webm")

# ... normal env.step() loop, or any pose changes ...

env.stop_recording()

format also accepts "png"/"jpg" (frame sequences). "gif" currently captures real frames but doesn't reliably trigger a download yet — use "webm" for now if you need the file to actually save.

Installing

Using pip

Swift is designed to be controlled through the Robotics Toolbox for Python. By installing the toolbox through PyPI, swift is installed as a dependency

pip install roboticstoolbox-python

Otherwise, Swift can be install by

pip install swift-sim

Available options are:

nb provides the ability for Swift to be embedded within a Jupyter Notebook

Put the options in a comma-separated list like

pip install swift-sim[optionlist]

Swift requires Python 3.10 or later.

From GitHub

To install the latest version from GitHub

git clone https://github.com/jhavl/swift.git
cd swift
pip install -e .

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

swift_sim-2.0.0rc1.tar.gz (1.9 MB view details)

Uploaded Aug 18, 2026 Source

Built Distributions

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

swift_sim-2.0.0rc1-cp313-cp313-win_amd64.whl (2.4 MB view details)

Uploaded Aug 18, 2026 CPython 3.13Windows x86-64

swift_sim-2.0.0rc1-cp313-cp313-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl (2.8 MB view details)

Uploaded Aug 18, 2026 CPython 3.13manylinux: glibc 2.17+ ARM64manylinux: glibc 2.28+ ARM64

swift_sim-2.0.0rc1-cp313-cp313-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl (2.8 MB view details)

Uploaded Aug 18, 2026 CPython 3.13manylinux: glibc 2.28+ x86-64manylinux: glibc 2.5+ x86-64

swift_sim-2.0.0rc1-cp313-cp313-macosx_11_0_arm64.whl (2.4 MB view details)

Uploaded Aug 18, 2026 CPython 3.13macOS 11.0+ ARM64

swift_sim-2.0.0rc1-cp312-cp312-win_amd64.whl (2.4 MB view details)

Uploaded Aug 18, 2026 CPython 3.12Windows x86-64

swift_sim-2.0.0rc1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl (2.8 MB view details)

Uploaded Aug 18, 2026 CPython 3.12manylinux: glibc 2.17+ ARM64manylinux: glibc 2.28+ ARM64

swift_sim-2.0.0rc1-cp312-cp312-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl (2.8 MB view details)

Uploaded Aug 18, 2026 CPython 3.12manylinux: glibc 2.28+ x86-64manylinux: glibc 2.5+ x86-64

swift_sim-2.0.0rc1-cp312-cp312-macosx_11_0_arm64.whl (2.4 MB view details)

Uploaded Aug 18, 2026 CPython 3.12macOS 11.0+ ARM64

swift_sim-2.0.0rc1-cp311-cp311-win_amd64.whl (2.4 MB view details)

Uploaded Aug 18, 2026 CPython 3.11Windows x86-64

swift_sim-2.0.0rc1-cp311-cp311-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl (2.8 MB view details)

Uploaded Aug 18, 2026 CPython 3.11manylinux: glibc 2.17+ ARM64manylinux: glibc 2.28+ ARM64

swift_sim-2.0.0rc1-cp311-cp311-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl (2.8 MB view details)

Uploaded Aug 18, 2026 CPython 3.11manylinux: glibc 2.28+ x86-64manylinux: glibc 2.5+ x86-64

swift_sim-2.0.0rc1-cp311-cp311-macosx_11_0_arm64.whl (2.4 MB view details)

Uploaded Aug 18, 2026 CPython 3.11macOS 11.0+ ARM64

swift_sim-2.0.0rc1-cp310-cp310-win_amd64.whl (2.4 MB view details)

Uploaded Aug 18, 2026 CPython 3.10Windows x86-64

swift_sim-2.0.0rc1-cp310-cp310-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl (2.8 MB view details)

Uploaded Aug 18, 2026 CPython 3.10manylinux: glibc 2.17+ ARM64manylinux: glibc 2.28+ ARM64

swift_sim-2.0.0rc1-cp310-cp310-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl (2.8 MB view details)

Uploaded Aug 18, 2026 CPython 3.10manylinux: glibc 2.28+ x86-64manylinux: glibc 2.5+ x86-64

swift_sim-2.0.0rc1-cp310-cp310-macosx_11_0_arm64.whl (2.4 MB view details)

Uploaded Aug 18, 2026 CPython 3.10macOS 11.0+ ARM64

File details

Details for the file swift_sim-2.0.0rc1.tar.gz.

File metadata

Download URL: swift_sim-2.0.0rc1.tar.gz
Upload date: Aug 18, 2026
Size: 1.9 MB
Tags: Source
Uploaded using Trusted Publishing? Yes
Uploaded via: twine/7.0.0 CPython/3.13.14

File hashes

Hashes for swift_sim-2.0.0rc1.tar.gz
Algorithm	Hash digest
SHA256	`fd6f8c50456c34716f9fb82468ae1067c2460f6b2d247acc7d96920a8757c1cc`
MD5	`606b51190d28df2d552e15eefca08451`
BLAKE2b-256	`cedbe4eddbf7a6c3253e44fed236343c171d57354c4c2d42e4e0d3812a3b921c`

Algorithm	Hash digest
SHA256	`05323f1255a5a668bd2efa3d310b4651c39a2b1c1ea210032785af7e154ec6e8`
MD5	`d870470daf972863f6954766c7c5986b`
BLAKE2b-256	`7119a0c2403bc90db6400e4871ac0c2ca05c64185ad12f20b0ef2f7e978a764e`

Algorithm	Hash digest
SHA256	`a049975beddbaa6dd1acb7ea35354bcdf18bf4aaef27be0c634367036b49da8e`
MD5	`028f7a46bba342676df72fd9e05d7516`
BLAKE2b-256	`4a553864449ca637e8c1c11952a907d6fd04b8a86dff0884341181bf631db44b`

Algorithm	Hash digest
SHA256	`1b9df960c5a0acf1ce78dac6fc489193249990e31363cac57880d1361f6c343f`
MD5	`90fc529103652da4267cd38924c8f3c4`
BLAKE2b-256	`42691f0bed0d0186ae2367fda9dfbaf6f2554e6fe60e77c5bf14ccd82a2e8dc2`

Algorithm	Hash digest
SHA256	`1764d3ce75eef049a179d87d559619813a7b2af02e360fa2f3b4042973ce32ca`
MD5	`3a604592cca54766e6d21487cf5fe440`
BLAKE2b-256	`7922b859c11e5aa5d2b05bbd92b1003d1d8ccc79720e3cc37123a072c8fba87a`

Algorithm	Hash digest
SHA256	`316a386e6f3e0baf3a726e304a97835a07ec93bd2010670d770b8a5c9db63862`
MD5	`5a81489698e78eb0cadc5e241934e186`
BLAKE2b-256	`8ed856abffa3775188cbd8d8bb4eea3a1e68c8b13a5c6c307ec9391778832636`

Algorithm	Hash digest
SHA256	`feee3ab3f84d330ebef93cf6c88dc1fe8b1904b9eaedae638c218fb7a87ab407`
MD5	`962bf37118279b3aa30b25924889be0d`
BLAKE2b-256	`39bbc164aaf6bf375a039f49b84863633dd3226bc38d97b0ee3dc83452a60f48`

Algorithm	Hash digest
SHA256	`f1aa0a603c05386d8f09ba6ff8e4f0c16a9b89dbcebc1aa12afab31e1c8df342`
MD5	`e9a6e4a958ea4a998370fc53eaeeec37`
BLAKE2b-256	`acaf227a58c17a558ea58496e12133ec926f580f59b10ab6ebf522b17fe96512`

Algorithm	Hash digest
SHA256	`0d91f24586d2456ab009edcc9b5b19227025910acfb7438ca3c4b26e222e91bb`
MD5	`e1deff16e1ff37dfb00b4bed11f0ed8e`
BLAKE2b-256	`e24d48a75dfca8794d90f489083788f25e677dff35ba20549e66aa74b1cc8721`

Algorithm	Hash digest
SHA256	`1b051d360bc4d0b822b527d3c0b4d880ed66ac84f194c5a0d19c2573af43a3ef`
MD5	`00486f632826438e029daecac8b9e1ec`
BLAKE2b-256	`db776d9069a167e8f4c30ac48327ea0a75d8a13e6d728e95935f20c2a2efb4a3`

Algorithm	Hash digest
SHA256	`4a9b4cd924a9049c83508a11096f88e2c634d593ae88334e1fc515754cc7d647`
MD5	`b068e37f5e62318371f1620620a83fc3`
BLAKE2b-256	`c8fcd40416a810deea43ed8b46297b763818d1cb95e6386f0df656d97ea83eed`

Algorithm	Hash digest
SHA256	`adc6a8537d7e470f7df9d550243a7f90cf373a6034004f44d9dc7dda82c03a44`
MD5	`92c9ca4df43145f0368db086dd51fc1b`
BLAKE2b-256	`e358ae6c2b5d12cbee69ca9c97ad70c920ca73939fddc4f9a34832c789b4069f`

Algorithm	Hash digest
SHA256	`fe025c9856a16e34af57c6d009397c66ea30298b7e53f00b120862e884c5990b`
MD5	`41ec9142102cdea6934be83ae5c70347`
BLAKE2b-256	`bbc46826511c00efb08684707d3064e48c2faf49dee61011b1ed2ef7ed751b11`

Algorithm	Hash digest
SHA256	`a69c2e7dea4af85d6eaa3d6855708c60193803497e7a7a08dbafa0d82488cc3b`
MD5	`e70cc60a753f85f9a4cfcf83b527ff4b`
BLAKE2b-256	`65923e373ad4bc9eeb9e7cb5060730e0d33773e0fef739f56619e5d0cf75f233`

Algorithm	Hash digest
SHA256	`9fa652a09c03ab78b79e7bad32c2361f661cb17efd28a15c832aecd9f72eb948`
MD5	`16ad43a971cf368f73797004654a51a3`
BLAKE2b-256	`63409f828aba19f5aeca503ff31ba8156a69472b9e286bcc2a022feb55bca893`

Algorithm	Hash digest
SHA256	`d4264564f0229602cf20bb14465f66e9832d72549dc6da1c1f2c397042ef28ab`
MD5	`1931a4afbac7196eb122a2e4fe0de6c4`
BLAKE2b-256	`5b51cb1581f109cc3d93ba19c2ee17a1e93f3c116086753c53f865830ed68c48`

Algorithm	Hash digest
SHA256	`eb0eaa8663d373ac20c100ebe858026128636b80dcc3ee9844a9ccacf2c955a9`
MD5	`4aabca38618017f6c130a8f7c78fcfcb`
BLAKE2b-256	`f0b011c492722c7411a88adf2d82792cbb55f396df978c9176a0bc3fd4cf66ec`

swift-sim 2.0.0rc1

Swift

What's new in 2.0

Examples

Render a box

Render a box, with sliders to move it

A 2-link arm, built from raw shapes (no robot model)

Render a Panda

A box driven by pure kinematics (no interaction)

Panda arm follows a target, positioned by sliders

Embed within a Jupyter Notebook

Playback controls

Recording video

Installing

Using pip

From GitHub

Project links

Key dates

2 maintainers

Credits

GitHub Statistics

License

Requires

Provides Extra

Tags

Classifiers

Download files

Source Distribution

Built Distributions

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes

Provenance

File details

File metadata

File hashes