README.md

metadata

license: apache-2.0
tags:
  - robotics
  - manipulation
  - simulation
  - pybullet

A2 Environment Assets

Assets for the A2 tabletop manipulation environment (PyBullet simulation).

Contents

3D Objects

• simplified_objects/: 66 simplified YCB objects for training
• unseen_objects/: 17 unseen objects for testing

Robot

• ur5e/: UR5e robot URDF and meshes
• workspace/: Workspace table and fixtures

Testing Cases

Pre-defined test scenarios for benchmark evaluation:

• testing_cases/grasp_testing_cases/: Grasp task test cases
  • seen/: 10 cases with training objects
  • unseen/: 5 cases with test objects
• testing_cases/place_testing_cases/: Place task test cases
  • seen/: 20 cases
  • unseen/: 10 cases
• testing_cases/pp_testing_cases/: Pick-and-place test cases
  • seen/: 8 cases
  • unseen/: 4 cases

Usage

Assets are automatically downloaded when using the A2 environment:

from lerobot.envs.a2 import A2Env

# Assets are downloaded to ~/.cache/a2_assets/
env = A2Env(task="grasp", num_objects=8)

Manual Download

from huggingface_hub import snapshot_download

# Download all assets
snapshot_download(
    repo_id="dgrachev/a2_assets",
    repo_type="dataset",
    local_dir="~/.cache/a2_assets"
)

Test Case Format

Each test case file contains:

• Line 1: Language instruction (e.g., "grasp a round object")
• Line 2: Target object indices (space-separated)
• Lines 3+: Object URDF path and 6D pose (x, y, z, rx, ry, rz)

Example (case00-round.txt):

grasp a round object
0 1 2
assets/simplified_objects/005.urdf 0.61 0.04 0.2 4.22 0.45 1.39
assets/simplified_objects/012.urdf 0.53 -0.11 0.2 6.15 3.49 5.06
...

Related Resources

• a2_pretrained - Pretrained A2 policy
• lerobot_policy_a2 - Policy package
• lerobot_grach0v - LeRobot fork with A2 environment

License

Apache-2.0