English | 简体中文

EBench: Elemental Mobile Manipulation Benchmark

From a single success rate to a multi-axis capability profile. Released by Shanghai AI Laboratory.

---

🔥 Recent Updates

💡 Interactive Analysis Tool

A new systematic diagnostic tool, gmp analyse, automatically produces a five-axis capability profile, such as atomic skills and manipulation precision, together with a four-dimensional generalization analysis after evaluation. Users can interactively select evaluation results and compare them with four reference baselines, π0, π0.5, X-VLA, and InternVLA-A1, across multiple capability and generalization dimensions. Example →

Date	Highlight
2026-09	🔧 Minor EBench fixes and improvements — Adjusted the step limits for bottle and shop from 3,000 to 5,000, and refined the web interface of the online evaluation platform.
2026-09	🛠️ OpenPI baseline update — Updated π0 / π0.5 configs. Please use the updated baseline code for evaluation and reproducing results. See the update notes.
2026-07	🛠️ Evaluation toolkit — genmanip-client brings submission, monitoring, action/state plots, and interactive episode visualization together in the gmp CLI.
2026-06	🚀 Public release — EBench, reference baselines, training data, and held-out online evaluation are now available.

What is EBench?

EBench is an indoor VLA manipulation benchmark built on NVIDIA Isaac Sim. Instead of compressing a model's behaviour into a single overall success rate, it produces a multi-axis capability profile that exposes what a model is good at — and where it overfits.

This repository is the project entry point. It hosts reference baselines and convenience scripts; the simulation runtime, the gmp CLI, and the datasets each live in their own repositories (linked in the badges above).

What makes EBench different

• Three manipulation regimes in one benchmark — covers long-horizon, dexterous & precise, and mobile manipulation, regimes most benchmarks address in isolation.
• 5-axis atomic diagnostic — every task is labelled by Scene · Atomic Skill · Horizon · Precision · Mobility, so a black-box score becomes a readable strength/weakness map.
• 4-axis generalization tests — controlled perturbations along Object · Background · Instruction · Mixed, attributing OOD drops to a specific axis.
• Strict train / test isolation — val_train and val_unseen are open for tuning; the held-out test (Test-Mini) drives the leaderboard, so numbers reflect real generalization rather than fitting to the eval distribution.

Two evaluation tracks are exposed: Specialist (Tabletop or Mobile Manip) and Generalist (both at once).

For the full methodology, task taxonomy, and per-axis rationale, see the project documentation.

Project layout

EBench is split across a small constellation of repositories. This repo is the front door:

Component	Where it lives	What it provides
EBench (this repo)	InternRobotics/EBench	Reference baselines, scripts, project entry point
GenManip	InternRobotics/GenManip	Isaac Sim evaluation server, task configs
genmanip-client	InternRobotics/genmanip-client	gmp CLI + EvalClient Python API
EBench-Assets	🤗 EBench-Assets	Scenes, objects, and task assets
EBench-Dataset	🤗 EBench-Dataset	Training trajectories (LeRobot format)
Docs site	internrobotics.github.io/EBench-doc	Setup, evaluation workflow, CLI reference
Online Challenge	internrobotics.shlab.org.cn/eval	Remote evaluation, leaderboard, diagnostic reports

EBench/
├── baselines/       # Reference policies (one sub-folder per baseline)
├── third_party/
│   └── genmanip-client/  # Pinned gmp CLI and Python client
├── scripts/         # Evaluation and analysis scripts
├── assets/          # Static assets used by this README
├── LICENSE
└── README.md

Quickstart

EBench runs as a client–server system. The server runs Isaac Sim; the client (gmp CLI) is a tiny package that drops into your model's Python environment.

# 1. Bring up the server  →  see Environment Setup
#    https://internrobotics.github.io/EBench-doc/getting-started/environment/

# 2. Clone EBench and its pinned dependencies
git clone --recursive https://github.com/InternRobotics/EBench.git
cd EBench

# If EBench was already cloned without --recursive:
git submodule update --init --recursive

# 3. Install the client in your model environment
pip install -e third_party/genmanip-client

# 4. Run an evaluation
gmp submit ebench/generalist/test --run_id my_first_run
gmp eval  -a r5a -g lift2 --worker_ids 0
gmp status

A full validation pass takes roughly 30 minutes on 8× RTX 4090. Detailed setup, asset download, and the complete gmp reference are in the docs site.

Evaluate with a coding agent

The repository includes five reusable agent skills for environment setup, policy integration, evaluation runs, debugging, and result analysis. They guide agents through the checked-out baseline and client implementations, including real-policy launch commands and result coverage checks.

To start, ask your agent: “Read skills/ebench-setup/SKILL.md and check whether this environment is ready to evaluate my model.” Then use ebench-evaluate to run it and ebench-analyze to interpret the results. See the skill catalog and usage guide for all workflows; automatic discovery depends on your agent's skill installation mechanism.

Tasks

26 task types across Long-Horizon, Pick-and-Place, and Dexterous & Precise, expanded with the four generalization axes and three splits into 794 evaluation task instances. Browse the video gallery at Task Showcase.

Baselines

Reference policies live under baselines/<name>/, each with its own README and a gmp eval-compatible entry point. EBench has been validated on π0, π0.5, X-VLA, and InternVLA-A1 — see the leaderboard for current standings and per-axis diagnostic reports.

Each baseline ships its upstream code as a third_party/ git submodule and layers an EBench-specific entry point on top. Initialize the submodules first:

git submodule update --init --recursive

openpi (π0 / π0.5)

Submodule lives at baselines/openpi/third_party/openpi; EBench-specific configs and the eval client are layered under baselines/openpi/{src,scripts}/. See baselines/openpi/README.md for the full walkthrough.

The post-trained OpenPI models on EBench are available at:

• π0.5 EBench Generalist
• π0 EBench Generalist

# After configuring paths/tokens in scripts/launch_pi_onlineeval.sh:
bash scripts/launch_pi_onlineeval.sh

X-VLA

# 1. Install deps
pip install -r baselines/X-VLA/requirements.txt

# 2. Run eval (each WORKER_ID is a separate inference worker)
MODEL_PATH=/path/to/xvla \
BASE_URL=https://internrobotics.shlab.org.cn/eval \
RUN_ID=my_first_run \
TOKEN=<your-token> \
WORKER_IDS=0 \
  bash scripts/run_xvla_eval.sh

InternVLA-A1

Submodule lives at baselines/InternVLA-A1/third_party/InternVLA-A1. See baselines/InternVLA-A1/README.md for the full walkthrough.

# 1. Install upstream deps from third_party/InternVLA-A1/README.md
# 2. Pull the checkpoint
huggingface-cli download hxma/EBench-Generalist-InternVLA-A1 \
  --repo-type model \
  --local-dir baselines/InternVLA-A1/checkpoints/EBench-Generalist-InternVLA-A1

# 3. Run eval
cd baselines/InternVLA-A1 && bash eval_pjsim.sh

To plug your own model in, follow the contract documented at Integrate Your Own Model.

Online challenge

The 24/7 evaluation platform at internrobotics.shlab.org.cn/eval runs every submission on the held-out Test-Mini split and produces an automatic diagnostic report (capability radar, validation→test transfer curve, generalization radar, task heatmap). Submission flow: see the Challenge page.

Citation

If you find this work helpful to your research, please consider citing it:

@article{gao2026ebench,
  title={EBench: Elemental Diagnosis of Generalist Mobile Manipulation Policies},
  author={Gao, Ning and Zheng, Jinliang and Gao, Xing and Ma, Haoxiang and Wang, Hanqing and Wang, Yukai and Chen, Jiantong and Chen, Zanxin and Zhang, Shujie and Jia, Mingda and others},
  journal={arXiv preprint arXiv:2606.18239},
  year={2026}
}

License

MIT — see LICENSE. Built on NVIDIA Isaac Sim, cuRobo, and the LeRobot data format. Issues and pull requests are welcome.