# RealMirror: an open source VLA platform and humanoid benchmark

> RealMirror is a research platform for vision-language-action models on humanoid robots, published under Apache-2.0 with simulation assets, model checkpoints and a benchmark spread across five manipulation tasks. The install path is demanding, since it expects NVIDIA Isaac Sim 5.0.0 and a patched LeRobot checkout, so the practical audience is research groups already inside that simulation stack.

**terminators2025/RealMirror** — RealMirror, a comprehensive, open-source embodied AI VLA platform.  

- Repository: https://github.com/terminators2025/RealMirror
- Website: https://terminators2025.github.io/RealMirror.github.io/
- Stars: 784 · Forks: 50
- Language: JavaScript
- License: Apache-2.0
- Published: 2026-10-20 · Updated: 2026-10-20 · Language: en
- Canonical page: https://hysenlabs.com/projects/terminators2025-realmirror

## RealMirror attacks the three costs that stall VLA research

The README names the obstacles it exists to remove: the high cost of data acquisition, the lack of a standardised benchmark, and the significant gap between simulation and the real world. That framing matters because each of those is usually solved by buying hardware, and the three together determine whether a small lab can publish comparable results at all.

The response is a platform that joins data collection, model training and inference so that end-to-end VLA work can proceed without a physical robot. The project description calls it an embodied AI VLA platform for humanoid robots, and the source tree backs that up with directories named simulation/, data_recording/, teleoperation/, inference/, evaluate/, controllers/, robots/ and tasks/. The intended users are therefore research teams rather than product engineers: people who need published baselines they can compare their own policy against.

The accompanying paper is on arXiv as 2509.14687, and the project page is hosted on GitHub Pages. The news entries record an acceptance to ICRA 2026, which tells you the work went through a peer review process, not that the code is stable.

## The benchmark ships assets and checkpoints as one downloadable set

The benchmark is distributed through two Hugging Face locations: zte-terminators/realmirror-asset for simulation assets and zte-terminators/realmirror-model-ckpt for pre-trained models. The download is stated to be approximately 13 GB, which includes both simulation assets and pre-trained models for all benchmark tasks, so treat the two repositories as a single delivery rather than picking one.

The layout is organised by task. Five scenarios are named in the tree: Task1_Kitchen_Cleanup, Task2_Cup_to_Cup_Transfer, Task3_Assembly_Line_Sorting, Task4_Can_Stacking and Task5_Air_Fryer_Manipulation. Under models, each task folder holds checkpoints for three model families, ACT, Diffusion and SmolVLA, which is what makes the set usable as a benchmark. A new policy can be measured against three existing approaches on identical tasks instead of against whatever the authors felt like comparing to.

The robot asset folder holds AgiBotA2 and A2_t2d0_flagship_opensource_urdf, and the project documentation recommends extracting the archive into a dedicated directory such as ~/RealMirror_Benchmark_data/. Getting this directory structure right matters, because the task configuration file for a run, for example Task1_Kitchen_Cleanup.json, resolves scene and robot paths from it.

## Installing RealMirror means patching a pinned LeRobot checkout

The install is not a pip install from a package index. The source is cloned and LeRobot arrives as a git submodule, because the platform depends on a modified version of that library:

```bash
git clone https://github.com/terminators2025/RealMirror.git
cd RealMirror
```

The submodule is fetched with Git LFS smudging disabled, then checked out to a specific commit and patched:

```bash
GIT_LFS_SKIP_SMUDGE=1 git submodule init
GIT_LFS_SKIP_SMUDGE=1 git submodule update
cd thirdparty/lerobot

# Checkout and apply patch of lerobot
git checkout fe88c5942cce222c5463350d999a463b9016cf8c
git apply ../../patches/lerobot_customizations.patch
```

Both editable installs then go into the Python environment that Isaac Sim ships with, rather than into a system Python:

```bash
$isaac_sim_dir/python.sh -m pip install -e .
```

The same command is run once inside thirdparty/lerobot for LeRobot and once in the repository root for RealMirror itself. Using $isaac_sim_dir/python.sh is the part teams get wrong: an install into a different interpreter produces import errors that look like missing dependencies later on.

## Sim2Real transfer is claimed for models trained only in simulation

The mechanism behind the real world claim is reconstruction rather than fine-tuning. The overview describes integrating generative models and 3D Gaussian Splatting to rebuild realistic environments and robot models, and reports zero-shot Sim2Real transfer in which models trained exclusively on simulation data perform tasks on a physical robot with no fine-tuning.

This is the part of the design worth scrutinising, because a simulator that renders its own training data can agree with itself and still disagree with the real world. The reconstruction step is what closes that gap, since the rendered scene is built to look like the physical environment and robot rather than to look like a generic simulation asset.

The evaluation framework, official model weights and robot and scene assets were released together on 2025-10-15, which is the point at which the project moved from paper to something a third party could download and run. Inference code and evaluation code arrived before the training code in the published plan, so a reader today has all three.

## Seed2Scale trades simulation speed for synthetic trajectories

The most recent piece of work is Seed2Scale, an automatic large-scale upper limb trajectory generation scheme built on the platform, described in arXiv paper 2603.08260 and announced on 2026-03-09. Its value to a user is a mode of generating training trajectories without a human in the loop.

That is the real alternative in this repository, and the difference in approach is concrete. The teleoperation and data acquisition toolchain released on 2025-10-30, together with the MirrorLimb work of 2025-11-12 that added gesture teleoperation and control through PICO devices, produces demonstrations from a person driving the robot. Seed2Scale produces trajectories automatically from a generative model, at a scale a single operator cannot reach. The trade-off is direct: teleoperation gives you human-chosen behaviour with a matching robot, while automatic generation gives volume and consistency at the cost of the demonstrator's intent.

The Seed2Scale scheme is still being extended. A sample dataset for automatic trajectory generation, together with visual physics playback tools and quantitative evaluation results, was marked as coming soon, and the opensource plan lists Seed2Scale example assets, example datasets and the playback tools as delivered before 2026-04-25.

## The hardware table decides who can realistically try this

The tested configuration is listed as a single row: Ubuntu 22.04, NVIDIA Isaac Sim 5.0.0, and a GeForce RTX 5090 or 5880 GPU with driver 575.64.03 and CUDA 12.9. There is no CPU-only path, no Apple Silicon path and no Windows path anywhere in the documentation.

The Python side is narrower still. setup.py declares python_requires='>=3.11' and the classifiers name Isaac Sim :: 5.0.0, so the supported range is Isaac Sim 5.0.0 exactly rather than a family of versions. The dependency list pins several packages tightly, including transformers==4.52.4, gymnasium==0.29.1, draccus==0.10.0 and grpcio==1.80.0, and that list is read directly from requirements.txt at install time, so a resolver conflict in your own environment becomes a RealMirror install problem.

One more constraint deserves attention. The install edits a submodule in place, at a fixed commit, and applies a patch file to it. If you pull new upstream commits into thirdparty/lerobot, the patch may stop applying, and the documentation does not describe how to resolve that.

## Licence is permissive, but release packaging is still research shaped

RealMirror is published under Apache-2.0, and the source headers carry the standard Apache License 2.0 text with a 2025 ZTE Corporation copyright. The permissive grant matters for commercial and academic reuse, and the licence file sits at the root of the repository. It covers the code; it does not tell you anything about the robot assets or the model checkpoints hosted on Hugging Face, which sit in separate repositories under the zte-terminators organisation and carry their own terms.

Release discipline is the weaker side. The repository has no GitHub releases, so there is no version history to roll back to, and the versioning visible in setup.py is version='0.1.0' for the package. That is consistent with a research platform rather than a shipped product, and it means upgrades arrive as commits on main rather than as tagged packages.

The last push was on 2026-05-18, and the repository is not archived. Since that date the project has shipped no tagged release, and the Seed2Scale sample dataset was still described as coming soon. Plan for manual updates rather than expecting a release train.

## Conclusion

Adopt RealMirror if your group already runs NVIDIA Isaac Sim 5.0.0 on Ubuntu 22.04 with RTX 5090 or 5880 hardware and wants a published benchmark with baseline checkpoints for ACT, Diffusion and SmolVLA instead of building assets from scratch. Skip it if your team has no GPU budget for that class of hardware, if you need a stable released package rather than a research tree with no GitHub releases, or if you must run on Windows or macOS. Before anything else, confirm that the thirdparty/lerobot submodule still checks out at commit fe88c5942cce222c5463350d999a463b9016cf8c and that patches/lerobot_customizations.patch applies cleanly to it, because the install instructions depend on both.

## FAQ

### What does RealMirror require to run on a workstation?

The tested setup is Ubuntu 22.04 with NVIDIA Isaac Sim 5.0.0 and a GeForce RTX 5090 or 5880 GPU using driver 575.64.03 and CUDA 12.9. The package also declares python_requires='>=3.11'.

### How do I get the RealMirror benchmark data and model checkpoints?

Both come from Hugging Face: zte-terminators/realmirror-asset holds the simulation assets and zte-terminators/realmirror-model-ckpt holds the pre-trained models. Together the download is approximately 13 GB and covers all five benchmark tasks with ACT, Diffusion and SmolVLA checkpoints.

### Why does the RealMirror install checkout a specific LeRobot commit and apply a patch?

LeRobot is pulled in as a git submodule under thirdparty/lerobot, checked out to commit fe88c5942cce222c5463350d999a463b9016cf8c, and patched with patches/lerobot_customizations.patch. Both LeRobot and RealMirror are then installed in editable mode into the Isaac Sim Python environment using $isaac_sim_dir/python.sh -m pip install -e .

### Can RealMirror models be transferred to a physical robot without fine-tuning?

The project reports zero-shot Sim2Real transfer for models trained exclusively on simulation data, using generative models and 3D Gaussian Splatting to reconstruct realistic environments and robot models. The approach is described in arXiv paper 2509.14687.

### What is Seed2Scale in the RealMirror project?

Seed2Scale is an automatic large-scale upper limb trajectory generation scheme built on the RealMirror platform, described in arXiv paper 2603.08260. It generates training trajectories without a human teleoperator, in contrast to the teleoperation and data acquisition toolchain released on 2025-10-30.

## Sources

- [Issues](https://github.com/terminators2025/RealMirror/issues)
- [License: Apache-2.0](https://github.com/terminators2025/RealMirror/blob/main/LICENSE)
- [Project website](https://terminators2025.github.io/RealMirror.github.io/)
- [README](https://github.com/terminators2025/RealMirror/blob/main/README.md)
- [terminators2025/RealMirror on GitHub](https://github.com/terminators2025/RealMirror)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/terminators2025-realmirror