echo-world 0.0.5

Preview — type a sentence, get a robot solving it in MuJoCo. Backend launching soon.

echo_world

🚧 PREVIEW RELEASE — backend launching soon. This package reserves the name and previews the API surface. The kernel at api.echo.dev is not live yet. Follow @EchoRobots on X for the launch.

The thin client for Echo — the kernel for the physical world.

pip install echo_world

.echo is one file format that carries an entire physical world: the CAD recipe, the kinematics, the materials, the affordances, the solver state, the trajectories. The Echo kernel reads it back into a live world that exports cleanly to STL, MJCF, URDF, glTF, USD, OBJ, STEP. Every consumer projection comes from the same source.

echo-world is the thin Python client. The compute lives at api.echo.dev. You don't install MuJoCo, you don't need a GPU, you don't manage meshes — you author intent, the kernel does the rest.

Three verbs, one file format

import echo

# 1. CREATE an asset — a thing
beaker = echo.create("a 100mL pyrex beaker")
arm    = echo.create("a 6-DOF servo arm with a parallel gripper")
arm.save("adam_arm.echo")          # ~12 KB

# 2. COMPOSE a world — assets + scene + task
world = echo.compose(
    things  = [beaker, arm],
    layout  = "the arm sits on a wooden table, beaker is to its right",
    task    = "pick up the beaker",
)
world.save("scene.echo")           # ~18 KB

# 3. SOLVE a world — run the task, get a trajectory
result = echo.solve(world)
result.video_url                   # https://echo.dev/v/abc.mp4
result.trajectory.save("run.echo") # replayable; another `.echo` file

Or in one line, when the prompt fully describes the scene:

result = echo.solve("Franka picks up a red ball off a wooden table")
print(result.video_url)

What .echo carries

Geometry	atoms with primitive types (Box, Cylinder, Mesh, Revolve, Sweep, …) and ops (Hole, Hollow, Fillet, …)
Kinematics	typed connections — bonds (HINGE, SLIDE, BALL, WELD), edges (curves)
Affordances	semantic tags on particles, faces, atoms (GRASP_FACE, PLACEMENT_TILE, HANDLE, …)
Materials	direct refs (PLA, ABS, glass, steel) — not free-form strings
State	atom poses, joint qpos, sim timestamp
Trajectories	replay buffers as state-deltas — million-step episodes in a few MB

Everything else — meshes, kinematic chains, convex hulls, collision patches, spatial relations — the kernel re-derives at load time. That's why a robot manifest is ~10 KB instead of ~100 MB.

Project to any consumer format

m = echo.load("scene.echo")

m.export.stl()        # → STL bytes (slicer, 3D printer)
m.export.mjcf()       # → MJCF XML (MuJoCo)
m.export.urdf()       # → URDF + sibling meshes (ROS)
m.export.gltf()       # → glTF 2.0 (web, Three.js)
m.export.usd()        # → USD (Isaac Sim, Omniverse)
m.export.obj()        # → OBJ (Blender, Maya)
m.export.step()       # → STEP (CAD tools)

Same source, every export. Edit the recipe → every projection updates.

Query a manifest

arm = echo.load("adam_arm.echo")

arm.atom("j2").pose                          # typed Pose
arm.atom("j2").particles_by_tag("BEARING_SEAT")
arm.bond("j1_to_j2").qpos = 0.5
arm.sync()
arm.atom("j6").world_pose.position           # follows the joint update

arm.particles_by_tag("GRASP_CONTACT")        # all grasp surfaces
arm.affordances()                            # all named affordances

No string parsing, no XML grep. Typed attributes on a Pydantic graph.

Solve from the shell

echo_world create "a 6-DOF servo arm with parallel gripper" -o arm.echo
echo_world compose --thing arm.echo --task "pick up the beaker"  -o scene.echo
echo_world solve scene.echo

# or one-shot
echo_world solve "Franka picks up a red ball off a wooden table"
#   video : https://echo.dev/v/abc123.mp4
#   .echo : https://echo.dev/m/abc123.echo

Get an API key

Backend is in development. Follow @EchoRobots on X for the launch — keys + free-tier signup will be announced there.

export ECHO_API_KEY=ek-...

Or pass it explicitly:

echo.Client(api_key="ek-...").solve(...)

Why a hosted backend

The SDK is ~200 lines of HTTP. The kernel is on Echo's servers — that means the same solver behavior for everyone, no local install of MuJoCo / Isaac / OCCT, no GPU requirement, and every fix lands for everyone the same minute. You author intent on your machine; the kernel runs in ours.

Status

v0.0.5 — preview. The package reserves the name and shows the API surface. The backend at api.echo.dev is not live yet — calls will fail with a friendly preview message until the kernel launches.

Follow @EchoRobots on X for the launch date and the v0.1.0 release.

License

MIT.