Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Eden

Robot-learning framework on the Genesis simulator: one typed config per task, GPU-parallel training, then deployment to real robots.

Reviewed by ReviewRamble editors · Updated

View on GitHub Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

3.8/5

Eden is a serious and unusually well-documented robot-learning framework, but at version 0.1.0 it suits labs that can absorb breaking changes, not teams that need stability.

What is Eden?

Eden (Embodied-AI Development Environments Nexus) is a Python framework for robot learning, not a curated list. It sits on the genesis-world physics simulator: you write a task as one typed Pydantic config covering the scene, the observations, actions and rewards, and the training runner, train it across thousands of parallel GPU environments, then run the policy in MuJoCo or on hardware with `eden deploy`.

It ships 47 robots and 11 composite rigs, 66 tasks (from locomotion to piano and calligraphy, plus ports of robosuite, LIBERO, SIMPLER and PolaRiS), 245 composable terms for rewards, observations, actions and other parts of a task, and a zoo of 21 published policies including π0.5 and GR00T. What it contains is code: it ships loaders for 12 or more public motion datasets rather than datasets of its own, and apart from Mochi, a solver for coupled rigid, cloth, soft-body and rod dynamics, the simulation comes from Genesis. Hardware backends cover the Unitree G1, Go2 and Go1, the LimX Oli, Franka arms and RoboTera and Wuji hands.

Is Eden worth it?

Eden is a young framework with an established lab's habits. The repository was created in May 2026, published v0.1.0 to PyPI as eden-world on 12 November 2026 and has 276 stars, yet its docs cover safety, timing and licensing in more detail than many mature robotics stacks. Two commits on 17 November were labelled breaking, and `main` pins an unreleased Genesis commit, so code written against today's docs may need changes within weeks.

The design follows the manager-based pattern of NVIDIA's IsaacLab and of mjlab, both credited in the README, with Genesis underneath instead of Isaac Sim. The appeal is breadth in one package: training with rsl-rl, rl_games, skrl or Stable-Baselines3 across GPUs, retargeting human motion onto robots, teleoperation through GELLO, Meta Quest, PICO or Wuji gloves, recording to HDF5 or LeRobot format, and export to ONNX or TorchScript. Linux with an NVIDIA GPU is the primary platform; macOS and Windows are best-effort, and Python must be 3.12 or 3.13.

Deployment is where the code moves real machines, and the defaults are cautious. A robot boots in STOP and moves only after an operator presses INIT and then START; with no keyboard or remote it holds still. The docs still tell you to keep the e-stop within reach, hang legged robots from a hoist for first runs and start at 0.2 of the trained gains. One documented gap matters: on the G1, Go2 and Oli a stalled control thread keeps re-sending the last command, so the robot's own watchdog never fires and stopping the process or pressing the e-stop is the only way down. `eden deploy` accepts rsl_rl checkpoints only.

Licences need a check before commercial use. Eden itself is Apache 2.0, but the optional GraspGen extra is non-commercial, SMPL body models are licence-gated, and the Wuji glove SDK sends anonymous usage events unless `WUJI_TELEMETRY_DISABLE=1` is set, which Eden does not set for you. The repository also ships an AGENTS.md and Claude Code skills for writing terms and porting mjlab tasks, so a coding agent can work in it under the same rules as human contributors.

Where it earns its place

  • A university humanoid lab with a Unitree G1

    Train a walking policy and test it on the real robot

    Train `velocity_tracking_flat` across 4,096 environments, check it in MuJoCo with `--mujoco --menagerie`, then deploy at `--deploy_kp_kd_scale 0.2` with the robot on a gantry and the Holosoma policy as the safety slot. The software is free; the costs are GPU time and a supervised session.

  • A manipulation researcher comparing against published baselines

    Run LIBERO, robosuite or SIMPLER tasks without three separate setups

    Eden reimplements those suites as its own tasks and runs zoo policies with `eden inference --policy <name>`. The README describes them as reimplementations, not the original code, which is worth stating in any paper that reports the numbers.

  • A robotics startup building a dexterous-hand demo

    Turn human hand motion into training data for a Wuji or XHand

    Use the retargeting extras (`dex-retargeter`, `pose`) to map motion onto the hand, record teleoperated demonstrations in LeRobot format and train. GVHMR checkpoints and SMPL models are separate downloads, and GraspGen is non-commercial.

Before you commit

Setup
`pip install eden-world` (not `eden`, which is an unrelated package), then `eden list` and `eden train --task reacher`. Needs Python 3.12 or 3.13; the first run compiles Genesis kernels, which takes a while. The examples need a clone and `uv sync`.
Integrations
genesis-world, MuJoCo for sim-to-sim checks, rsl-rl, rl_games, skrl and Stable-Baselines3; HDF5 and LeRobot datasets; ONNX and TorchScript export; GELLO, Meta Quest, PICO, Wuji and SONIC teleoperation; Unitree, LimX, Franka, RoboTera and Wuji hardware.
Data and privacy
Runs on your own machines. The optional Wuji SDK sends anonymous usage events to a third-party analytics host unless `WUJI_TELEMETRY_DISABLE=1` is set. Code is Apache 2.0; GraspGen is non-commercial and SMPL models are licence-gated.
Cost at scale
Free. Training at scale needs NVIDIA GPUs, with multi-GPU through DDP; CPU mode is enough to build tasks and step a few environments. Robots, safety rigs and operator time are the real costs.
Limits
Nyx path tracing has no macOS build, Gaussian-splat rendering needs CUDA, `eden deploy` supports rsl_rl checkpoints only, and composite rigs with parts that have no hardware backend (KukaSharpa, the G1's Inspire hands) are refused on hardware.

What’s good

  • One typed config per task, with variants as CLI flags
  • 47 robots, 66 tasks and a 21-policy zoo
  • Cautious, thoroughly documented hardware deployment
  • Training, teleoperation, retargeting and export in one package

What’s not

  • Version 0.1.0 with frequent breaking commits
  • Main branch pins an unreleased Genesis commit
  • A stalled control thread keeps sending the last command
  • Non-commercial and telemetry caveats in some extras

Best for

Robotics researchers and lab engineers on Linux with NVIDIA GPUs who want one framework from simulated training to a real robot.

Eden: common questions

Is Eden worth it?
Eden is a serious and unusually well-documented robot-learning framework, but at version 0.1.0 it suits labs that can absorb breaking changes, not teams that need stability. We rate it 3.8 out of 5.
How much does Eden cost?
Free under Apache 2.0. Some optional extras carry other terms: GraspGen is under NVIDIA's non-commercial licence, and SMPL body models are licence-gated.
What are the downsides of Eden?
Version 0.1.0 with frequent breaking commits; main branch pins an unreleased Genesis commit; a stalled control thread keeps sending the last command; non-commercial and telemetry caveats in some extras.
Who is Eden best for?
Robotics researchers and lab engineers on Linux with NVIDIA GPUs who want one framework from simulated training to a real robot.
How much does Eden cost for a team or in production?
Free. Training at scale needs NVIDIA GPUs, with multi-GPU through DDP; CPU mode is enough to build tasks and step a few environments. Robots, safety rigs and operator time are the real costs.
Is Eden safe for company data?
Runs on your own machines. The optional Wuji SDK sends anonymous usage events to a third-party analytics host unless `WUJI_TELEMETRY_DISABLE=1` is set. Code is Apache 2.0; GraspGen is non-commercial and SMPL models are licence-gated.

Answers checked Nov 18, 2026.