UniLab

A Heterogeneous Architecture for Robot RL Beyond GPU-Dominant Paradigms.

Languages: English | 简体中文

🎉 🎉 UniLab has been accepted to CoRL 2026! 🎉 🎉

One task-authoring surface for locomotion, manipulation, and motion tracking.

UniLab is a complete, configurable product for robot reinforcement learning. Describe a task with Hydra, assemble it from manager terms, select a physics backend, and train or evaluate through one CLI. The same task-facing contract connects CPU, GPU, and external-worker simulation to the learner runtime.

Physics adapters are provided by the independent unisim-core package. RL algorithms and their runners are provided by unilab-rl (Python namespace uni_rl). UniLab keeps the user-facing task, environment, configuration, and experiment workflow together.

New to UniLab? Start with First success. Already have a task? Jump to Train and evaluate and change only --sim to try another backend when a matching task owner is available.

Highlights

┌──────────────────────────────────────┐     Same task contract    ┌──────────────────────────────────────┐
│                                      │ ────────────────────────▶ │       Run it where you need          │
│       Define the task once           │                           │   MuJoCo · Motrix · MJWarp · Drake   │
│       Hydra · Managers · NumPy       │                           │    Genesis · IsaacGym · IsaacSim     │
│     Terms · rewards · commands       │                           │   CUDA · ROCm · macOS · MPS · XPU    │
│                                      │                           │         train · eval                 │
└──────────────────────────────────────┘                           └──────────────────────────────────────┘

UniLab's core idea is simple: define task semantics once as reusable configuration, then change the simulator, hardware, or learner without rewriting the task's environment lifecycle.

• Configure, don't code. Actions, observations, rewards, terminations, events, commands, curricula, and metrics are manager terms assembled in Hydra owner YAML. Variants built from existing terms need no new environment class — often no Python code at all.
• Change the backend, keep the workflow. Current and future simulators share the public SimBackend contract. Choose a backend with --sim; the same task authoring and train/eval workflow remains in place while the owner YAML keeps backend-specific details explicit.
• Scale across the hardware you have. CPU-parallel or external-worker simulation feeds accelerator learners through the injected env contract and async runtime. Algorithms and runners are supplied by the unified package ecosystem instead of being tied to one simulator.

Getting started

The supported source workflow uses uv.

curl -LsSf https://astral.sh/uv/install.sh | sh
git clone https://github.com/unilabsim/UniLab.git
cd UniLab

# Fastest path to the first Motrix demo.
make setup-motrix

# Full local setup (MuJoCo + Motrix):
# make setup

# Optional platform/backend paths:
# make sync-rocm       # AMD GPU
# make sync-xpu        # Intel GPU
# make setup-drake     # Drake + native batch extension

The mujoco extra compiles a native extension and may require the platform compiler and Python development headers. See the installation guide for platform-specific setup, optional backends, and external worker runtimes.

First success: run a demo

# Downloads the checkpoint and assets from Hugging Face on first run.
uv run demo dance

Available presets are teaser, dance, wallflip, boxtracking, locomani, and inhandgrasp. Use uv run demo --help for device and refresh options. The quick demo guide explains rendering modes and server/macOS differences.

Train and evaluate

# Train and replay a task with Motrix.
uv run train --algo ppo --task go2_joystick_flat --sim motrix
uv run eval --algo ppo --task go2_joystick_flat --sim motrix --load-run -1

# Switch only the simulator for the same task.
uv run train --algo ppo --task go2_joystick_flat --sim mujoco

# Or use the same workflow with an off-policy learner.
uv run train --algo sac --task g1_walk_flat --sim mujoco
uv run train --algo flashsac --task g1_walk_flat --sim mujoco

# Headless video export.
uv run eval --algo ppo --task go2_joystick_flat --sim motrix \
  --load-run -1 --render-mode record

Route-defining choices are always visible:

--algo + --task + --sim  →  Hydra owner YAML  →  registered environment

Use normal Hydra overrides after those flags:

uv run train --algo ppo --task go2_joystick_flat --sim motrix \
  algo.max_iterations=1 algo.num_envs=16 training.no_play=true

Do not override training.sim_backend to switch engines. It is the identity field supplied by the selected owner YAML. Find resume, W&B, playback, and the full command matrix in the training guide.

Manager-based configuration

UniLab wraps a community-familiar manager API with Hydra composition and a NumPy runtime. A task owner can select and parameterize terms declaratively:

env:
  observations:
    policy:
      terms:
        joint_pos:
          func: unilab.envs.mdp.joint_pos_rel
        command:
          func: unilab.envs.mdp.generated_commands
          params:
            command_name: twist
  actions:
    joint_pos:
      _target_: unilab.envs.mdp.JointPositionActionCfg
      entity_name: robot
      scale: 0.25
reward:
  tracking_lin_vel:
    func: unilab.tasks.locomotion.common.manager_terms.track_lin_vel_xy_exp
    weight: 1.0

This makes common task edits a config change: compose or disable a term, tune its parameters, and reuse it across robots and backends without writing a new environment class. The API follows the pinned mjlab manager semantics where the contracts are shared, but it is a UniLab product with NumPy, Hydra, and NpEnvState semantics. See the Manager-Based API guide for the complete contract and known differences.

Physics backends

Current backends are available through unisim-core and the same UniLab route; the contract is designed to grow as new adapters land:

mujoco · motrix · mjwarp · drake · genesis · isaacgym · isaacsim

Choose one backend setup (combine extras when needed):

uv sync --extra mujoco
# uv sync --extra mujoco --extra motrix
# uv sync --extra mujoco --extra mjwarp
# uv sync --extra genesis
# make setup-drake

IsaacGym and IsaacSim use dedicated external worker environments. Backend installation details, rendering behavior, and the evidence-based task support matrix live in the backend guide and support matrix.

Ecosystem

UniLab is designed to be the shared product surface for robot-specific repositories. Current downstream examples include MicroDuck RL and EngineAI RL. They can ship robot recipes independently while consuming the same task, backend, and RL contracts.

Documentation

• Documentation index
• Unified CLI reference
• Task and manager architecture
• Sim-to-sim deployment
• Algorithm extension recipe
• Architecture decisions

For development and contribution workflows, see the contributing guide.

Community

Add the UniLab assistant on WeChat to join the community.

Citation

@article{jia2026unilab,
  title         = {UniLab: A Heterogeneous Architecture for Robot RL Beyond GPU-Dominant Paradigms},
  author        = {Jia, Yufei and Cao, Zhanxiang and Yu, Mingrui and Zhang, Heng and Chen, Shenyu and Jiang, Dixuan and Li, Meng and Li, Xiaofan and Liu, Yiyang and Wu, Junzhe and Li, Zheng and Fang, XiLin and Cui, Tingyu and Fu, Shengcheng and Li, Haoyang and Wang, Anqi and Wang, Zifan and Zhu, Dongjie and Cao, Chenyu and Huang, Zhenbiao and Zheng, Ziang and Lu, Jie and Ma, Xin and Wei, Zhengyang and Zhao, Xiang and Zhan, Tianyue and He, Ye and Chen, Yuxiang and Jiang, Yizhou and Li, Yue and Ge, Haizhou and Dong, Yuhang and Jia, Fan and Zhang, Ziheng and Zhang, Meng and Deng, Xiwa and Chen, Zhixing and Shao, Hanyang and Dong, Chenxin and Li, Yixuan and Chen, Yizhi and Chen, Bokui and Zhang, Kaifeng and Cui, Hanqing and Qin, Yusen and Huang, Ruqi and Han, Lei and Wang, Tiancai and Li, Xiang and Gao, Yue and Zhou, Guyue},
  journal       = {arXiv preprint arXiv:2605.30313},
  year          = {2026},
  url           = {https://arxiv.org/abs/2605.30313}
}

UniLab is released under the Apache License 2.0. See the independent UniSim and UniLab RL repositories for their own release and citation information.

Acknowledgments

UniLab would not exist without the excellent work of the Isaac Lab team and the mjlab developers and contributors. Isaac Lab's manager-based API design and abstractions, together with mjlab's clear, lightweight reference implementation, helped shape UniLab's Hydra and NumPy task authoring experience. We sincerely thank both communities for sharing their work and ideas.