Skip to content

Releases

Eden 0.1.0

Choose a tag to compare

@Kashu7100 Kashu7100 released this 12 Nov 11:32
527dd12

Eden 0.1.0 is the first public release of Eden (Embodied-AI Development Environments Nexus), published on PyPI as eden-world under Apache-2.0. Eden is a robot-learning framework built on Genesis (genesis-world==1.4.6): vectorized environments for locomotion, manipulation, dexterous and motion-imitation tasks, configured declaratively with Pydantic and composed from pluggable terms. This is an alpha release; APIs may change before 1.0.

Highlights

Write a robot-learning task as one typed config, train it across thousands of GPU-parallel environments, then run the same policy in MuJoCo or on a real robot.

  • One typed config per task. Scene, MDP and learning runner live in one Pydantic object; task variants become CLI flags.
  • 66 tasks, from locomotion, parkour and piano to calligraphy and painting, plus ports of robosuite, LIBERO, SIMPLER and PolaRiS.
  • 47 robots and 11 composite rigs: humanoids (G1, Oli, SMPL), quadrupeds (Go1, Go2), arms (Franka, UR5e/UR10e, xArm7, iiwa14, WidowX, YAM) and dexterous hands.
  • 245 composable MDP terms, searchable with eden terms --grep <text>; any decorated function becomes a new term.
  • Multi-physics: every Genesis solver, plus the Mochi solver for coupled rigid, cloth, soft-body and rod dynamics.
  • Train and deploy: rsl_rl, rl_games, skrl or Stable-Baselines3 on multi-GPU DDP, then eden deploy to MuJoCo or hardware, or eden export to ONNX or TorchScript.
  • 21-policy zoo, including SONIC, BeyondMimic, π0.5 and GR00T, run with eden inference --policy <name>.
  • Built for coding agents: strict configs, error messages that name the fix, and a shipped AGENTS.md.

What's included

Every item below ships in 0.1.0; the right-hand column is where to start.

Area In 0.1.0 Start with
Tasks 66 tasks: velocity tracking, extreme parkour, motion imitation (BeyondMimic, GeneralMimic, ConTrack, AssistMimic), dexterous lift and reorient, piano, calligraphy, painting and whiteboard writing, Mochi cloth and cable tasks; 12 PolaRiS, 11 robosuite, 4 SIMPLER and the LIBERO ports eden list
Robots 47 robots and 11 composite rigs (for example DROID, G1 + Inspire, xArm7 + XHand1), and 950+ objects docs/guides/robots.md
MDP terms 14 action, 64 observation, 46 reward, 34 termination, 7 command, 32 event, 6 curriculum, 20 metric and 7 recorder terms, plus 15 noise and modifier terms eden terms --grep <text>
Physics Every Genesis solver; the Mochi solver for coupled rigid, cloth, soft-body and rod dynamics docs/guides/physics.md
Sensing and rendering RGB-D, lidar, tactile, IMU and audio; batched Gaussian-splat scenes; Nyx path tracing docs/guides/sensors-and-rendering.md
Learning rsl_rl (core), rl_games, skrl, Stable-Baselines3, CleanRL via gym.make_vec("Eden/<task>"), FlashSAC; multi-GPU DDP on one or many nodes eden train
Deployment Sim-to-sim against MuJoCo, real hardware, ONNX or TorchScript export with a policy_spec.json eden deploy, eden export
Policy zoo 21 frozen or server-backed policies, including SONIC, BeyondMimic, π0.5, GR00T, Octo and HOMIE eden inference --policy <name>
Motion data Human pose from video (GVHMR), retargeting (SOMA, OT, dex, FlashDex), dataset preprocessing eden pose, eden retarget, eden preprocess
Teleoperation and datasets GELLO, Meta Quest, PICO, Wuji Glove and SONIC teleop; scripted demos; HDF5 recording and LeRobot conversion eden teleop, eden datagen, eden lerobot
Tooling eden smoke build check, eden view, eden arrange, eden snap, a runnable examples catalog eden --help

Installation and requirements

Install the distribution eden-world; the import name and the command are both eden. Do not pip install eden: that PyPI name belongs to an unrelated project.

pip install eden-world
pip install "eden-world[vis]"       # with an optional extra

To run the examples or work on Eden itself, install from a clone:

git clone https://github.com/embodied-ai-nexus/Eden.git
cd Eden
uv sync --all-extras --all-groups   # every extra, including the git-only packages
Requirement Supported in 0.1.0
Python 3.12 and 3.13 (not yet run on 3.14)
Primary platform Linux x86_64 with an NVIDIA GPU
Best-effort platforms macOS on Apple silicon, Windows AMD64
Physics engine genesis-world==1.4.6
CPU-only Every command takes --cpu; fine for building tasks, not for training at scale

The core covers building tasks, rsl_rl training, inference, TorchScript export and the asset cache. Rendering, other RL libraries, retargeting, grasp sampling, teleoperation devices and ONNX are optional extras, listed in the installation guide.

Fixes since the initial commit

Two fixes landed on main after the initial release commit, and both are in 0.1.0.

PR Change Effect for users
#2 README images load from absolute raw.githubusercontent.com URLs The logo, teaser and gallery render on the PyPI project page
#1 mediapipe==0.10.14 gated to Python < 3.13; uv pinned to 0.13.7 in CI uv sync resolves again on uv 0.13+; on Python 3.13 the dex-retargeter-example extra installs without mediapipe

Known issues and limitations

0.1.0 is an alpha: configs, term parameters and CLI flags may change in a minor release.

  • Platforms. macOS and Windows are best-effort. Nyx path tracing and onnx-gpu have no macOS build, Gaussian-splat rendering needs CUDA, and the Wuji Glove SDK is Linux-only.
  • Python 3.14 is not supported yet (requires-python = ">=3.12,<3.14").
  • The PyPI package carries eden only. examples/, the tests and the docs live in the repository, so eden examples and the example scripts need a clone.
  • The video hand-retargeting example does not run on Python 3.13, because mediapipe==0.10.14 has no 3.13 wheels.
  • onnx and onnx-gpu exclude each other: both runtimes install into the same onnxruntime/ directory. Pick one.
  • Extras that need more than pip. soma-retargeter installs only with uv in a clone. metaquest and flashsac need git-only packages: in a clone, add --group metaquest-git or --group flashsac-git.
  • Checkpoints depend on the rsl-rl-lib version. Its 3.x and 4.x/5.x .pt layouts differ and both are admitted, so a checkpoint may not replay in another environment.
  • README links on PyPI. Images render on the PyPI page, but relative links (AGENTS.md, LICENSE, examples/) do not resolve there.

Links