Skip to content

What ships

What you get with pip install alloy-engine 0.1.0, and what more you can download: ready-made environments, robots and objects. Every name and count on this page was read from the 0.1.0 package.

import alloy
import alloy.envs
from alloy.api.catalog import objects
from alloy.cards.library import default_root

for name in sorted(alloy.envs.REGISTRY):
    print(name)
robots = sorted(p.name for p in (default_root() / "robots").glob("emb_e?_*")
                if not p.name.endswith(("_route", "_stiff")))
print(len(robots), "robots,", len(alloy.materials()), "materials,", len(objects()) >= 118)

ClothFold-v0
DrawerOpen-v0
GripperLift-v0
GripperLiftCamera-v0
PandaReach-v0
PegInsert-v0
PickCube-v0
PickPlace-v0
Pour-v0
9 robots, 90 materials, True

Ready-made environments

alloy.envs.make("<name>", num_envs=64) builds one (Train with RL). All nine give state observations; one adds a camera image.

Name Task Observations
GripperLift-v0 a two-finger gripper lifts a cube 8 cm and holds it state
GripperLiftCamera-v0 the same task state and a camera image
PandaReach-v0 a Franka Panda arm moves its hand to a random target state
PickCube-v0 a Franka Panda arm picks a 4 cm cube and carries it to a point in the air state
PickPlace-v0 a gantry gripper puts a scanned object in a bin state
PegInsert-v0 a gripper inserts a 10 mm peg 2 cm into a 10.5 mm hole state
DrawerOpen-v0 a gripper pulls a cabinet drawer 15 cm open state
Pour-v0 move a bowl under a stream of water and catch it state
ClothFold-v0 experimental: fold a towel along its diagonal; slow, needs a GPU state

Two more for legged robots: from alloy.envs.loco import go2_velocity, go2_rough (a Unitree Go2 following velocity commands on flat ground, and on rough terrain).

Robots

Nine robot files are in the package. Load one with scene.add_robot("panda", default_root() / "robots" / "emb_e2_panda" / "scene.xml") (Robots).

Folder Robot Kind Licence of the source files
emb_e1_fr3 Franka FR3 with the Franka Hand arm Apache-2.0
emb_e2_panda Franka Panda with the Franka Hand arm Apache-2.0
emb_e3_ur10e_2f85 Universal Robots UR10e with a Robotiq 2F-85 gripper arm BSD-3-Clause
emb_e4_fr3_bimanual two Franka FR3 arms two arms Apache-2.0
emb_e5_fr3_duo Franka Mobile FR3 Duo: a mobile base, a spine, two FR3 arms mobile, two arms Apache-2.0
emb_e6_g1 Unitree G1 with Dex3 hands humanoid BSD-3-Clause
emb_e7_h1_2 Unitree H1-2 with Inspire hands humanoid BSD-3-Clause
emb_e8_go2 Unitree Go2 quadruped BSD-3-Clause
emb_e9_leap LEAP Hand (right) hand MIT

Your own robot loads from an MJCF or URDF file the same way.

Objects

In the package: 118 object cards and 90 materials. objects() lists the cards on your disk; scene.add_object("mug", "<card name>", pos=(0, 0, 0.4)) places one.

On the Hugging Face Hub: 5,423 more object cards, each with its meshes and textures. Download by name:

pip install "alloy-engine[import]"
alloy fetch --list                 # what there is, by library and category
alloy fetch ycb_025_mug            # one object, into ~/.cache/alloy/imports
Library Objects What Licence
YCB 78 scanned household objects CC BY 4.0
Google Scanned Objects 1,033 scanned household objects CC BY 4.0
Objaverse 1,989 everyday objects in 240 categories CC BY 4.0 (1,965), CC0 1.0 (24)
Kitchen objects 1,900 kitchen objects and fixtures by Lightwheel, as packaged by RoboCasa CC BY 4.0
Gym-Anything parts 393 one card per rigid part of its articulated objects MIT
Alloy's own 30 household, kitchen, lab and snap-fit parts Apache-2.0 (25), CC0 1.0 (5)

No account is needed. The kitchen library is objects only: no whole kitchen scenes ship.

Keep in mind

  • Most object masses are estimates, not measurements: 5,180 of the 5,423 Hub cards are marked as estimates. Weigh the object or read its card before a result depends on its mass.
  • Friction comes from the material a card names, never from a measurement of that object.
  • Each card records its source's licence and credit line; check both before commercial use.
  • ClothFold-v0 is experimental and slow; training an arm to grasp does not work reliably yet (Train with RL).

Next