robox3d

Lightweight robot simulation in Python, with a viewer that runs in your browser.

robox3d wraps Box3D — the new 3D physics engine by Erin Catto (author of Box2D) — into a batteries-included robotics toolkit: URDF loading, position/torque control, F/T·IMU·LiDAR·contact sensors, and WebSocket pose streaming to a React Three Fiber viewer. Physics runs headless; the viewer is just a browser tab.

A bundled SO-ARM101 simulated by robox3d and teleoperated with joint sliders — everything above runs from pip install + one command, visualized in a plain browser tab.

pip install "robox3d[viz]"
python -m robox3d.demo so101   # SO-ARM101 with joint sliders — opens at http://localhost:8765

日本語のREADMEは README.ja.md にあります。

Why robox3d?

• Browser-native visualization — no OpenGL window, no X forwarding. The sim serves its own web viewer (HTTP + WebSocket on one port). Run headless on a server, watch from your laptop. Joint sliders in the viewer drive the robot live.
• Zero-friction install — a small C17 engine with no dependencies, shipped as prebuilt wheels.
• URDF in, physics out — links, joints, inertia, collision meshes (convex hull or CoACD decomposition), visual meshes and colors all handled.
• Robotics-grade control & sensing — spring position control with physical-unit gains (kp in N·m/rad, DC-calibrated), pseudo torque control, gravity-compensation feedforward, and F/T, IMU, LiDAR (batched raycasts), and contact sensors.
• Deterministic & fast — bit-exact reproducible across thread counts (validated), ~44,000 steps/s (≈180× real time) for a 6-DoF arm with active position control at 240 Hz × 4 substeps on a single desktop CPU core, including per-step Python-side target writes.
• Record & replay — pose recordings (.rbx) use the same wire format as live streaming; replay them into the same viewer.

Quickstart

import numpy as np
import robox3d

with robox3d.World() as world:   # Z-up, gravity (0, 0, -9.81)
    ground = world.create_body(kind="static")
    ground.add_box(half_extents=(10, 10, 0.1), offset=(0, 0, -0.1))

    link = world.create_body(position=(0, 0, 1))
    link.add_capsule((0, 0, 0), (0, 0, -0.5), radius=0.05)

    joint = world.create_revolute_joint(
        ground, link, anchor=(0, 0, 1), axis=(0, 1, 0),
        spring=(60.0, 1.0),      # spring position control (hertz, damping_ratio)
    )
    joint.target_angle = np.radians(45)

    for _ in range(240):
        world.step(1 / 240)      # substeps=4 (default)
    print(np.degrees(joint.angle))

Load a robot from URDF

robot = robox3d.load_urdf(world, "robot.urdf")   # fixed base, self-collision off
robot.enable_position_control(hertz=120.0)
robot.set_targets(q_des)                          # auto-clamped to URDF limits
q, qd = robot.positions(), robot.velocities()

An SO-ARM101 model (TheRobotStudio SO-101, Apache-2.0) is bundled:

robot = robox3d.load_urdf(world, robox3d.assets.so101())

Mesh collision defaults to a single convex hull per link. For finer shapes use mesh_mode="coacd" (pip install "robox3d[coacd]"; results are cached in ~/.cache/robox3d/).

Visualize in the browser

from robox3d.viz import VizServer

server = VizServer(world, robot=robot)  # needs: pip install "robox3d[viz]"
server.start()                          # call after all bodies are created
while running:
    world.step(dt)
    server.update()                     # stream poses + apply viewer commands

Open server.url (default http://127.0.0.1:8765) — the same port serves the viewer over HTTP and streams poses over WebSocket. Passing robot= adds joint sliders to the viewer. URDF <visual> elements (colors and meshes included) are displayed automatically, with a HUD toggle for collision shapes.

Record and replay (same format as the live stream):

from robox3d.viz import PoseRecorder
with PoseRecorder(world, "run.rbx") as rec:
    for _ in range(1000):
        world.step(dt)
        rec.update()

python -m robox3d.viz.record run.rbx   # replays to the viewer

Control and sensors

# Gains in physical units (kp: N·m/rad); steady-state stiffness is calibrated
robot.enable_position_control(kp=200.0)

# Torque control (gravity-compensation feedforward + spring feedback)
robot.enable_torque_control(disable_springs=False)
robot.set_torques(robot.gravity_compensation())   # call every step

# Sensors
ft = robox3d.FTSensor(mount_joint, frame="joint") # constraint force/torque
imu = robox3d.IMU(body)                           # specific force + gyro
lidar = robox3d.Lidar(world, body=base, num_rays=360, max_range=10)
touch = robox3d.ContactSensor(foot_body)          # contact + net normal force

For many joints/bodies in an RL or control loop, the batch API (RevoluteGroup / BodyGroup) turns per-step FFI traffic into O(1) calls:

group = robox3d.RevoluteGroup(joints)
group.set_targets(q_des)          # np.ndarray (n,)
q, qd = group.angles(), group.speeds()

Examples

Every example serves the browser viewer by default and keeps the scene moving until you Ctrl+C — open the printed URL to watch it live (real-time paced). Use --headless for a fast numeric run without the viewer, and --port to run several examples at once.

Scope and honest limitations

Box3D is a maximal-coordinate rigid-body engine (like game physics, unlike MuJoCo's generalized coordinates). robox3d validates and documents what that means in practice (validation report):

• Joint drift is negligible (0.003 mm over 20 s on a 6-link chain) and 1:100 mass ratios stay stable, but contact-rich scenes are the sweet spot — precise dynamics studies should cross-check against Pinocchio/MuJoCo.
• Revolute joint limits are capped at ±0.99π by the engine; wider URDF limits fall back to command clamping.
• substeps < 4 is rejected — the solver needs substepping for stiff chains.
• On chains where parallel hinges are bracketed by perpendicular ones (most arms), the engine's axis-alignment constraint leaks a small torque into the hinge, proportional to the joint constraint stiffness. robox3d's position control defaults to a tuning that keeps this below ~3° and exposes a knob (enable_position_control(constraint_hertz=...)) to trade pivot rigidity for sub-0.01 rad tracking. Full analysis: docs/spring-chain-investigation.md.
• Simulation is deterministic across thread counts; recordings are bit-stable.

Development

git clone --recursive https://github.com/neka-nat/robox3d
cd robox3d
uv sync                        # builds box3d + shim via scikit-build-core
uv run pytest                  # 56 tests
uv run python tools/build_viewer.py   # bundle the web viewer (needs pnpm)
uv run python examples/viz_arm.py

The Box3D version is pinned via the external/box3d git submodule. When the upstream API changes, re-run uv run python tools/gen_ffi.py and review the diff — the cffi bindings are generated from the C headers.

Architecture (details in docs/development-plan.md, Japanese):

Credits

• Box3D by Erin Catto (MIT)
• Bundled SO-ARM101 model by TheRobotStudio (Apache-2.0)

License

MIT