BeTTER: Diagnose the Illusion of Embodied Reasoning in Vision-Language-Action Models
Haiweng Xu, Sipeng Zheng, Hao Luo, Wanpeng Zhang, Zongqing Lu
ECCV 2026
Peking University, BeingBeyond
This repository is the official code release for BeTTER, the ECCV 2026 paper BeTTER: Diagnose the Illusion of Embodied Reasoning in Vision-Language-Action Models. It provides a benchmark and tooling stack for building and editing embodied reasoning tasks in Isaac Sim.

Recent Vision-Language-Action models report strong scores on standard robotic benchmarks, but those scores can hide weak semantic grounding and brittle sequential reasoning. BeTTER is designed to expose those failures by constructing targeted task variations while keeping low-level execution factors controlled.
This initial public release includes:
assets/task_templates/assets/services/retrieval_server/Large assets are distributed through the companion Hugging Face dataset repository: Seaman05/BeTTER-assets.
Currently published asset bundles include:
The published asset bundles should be unpacked with relative paths into the repository tree.
For local development, download the asset bundles from Hugging Face and either unpack them directly into the corresponding assets/ subdirectories or symlink those directories into your local checkout.
Because BeTTER is built on Isaac Sim 5.0.0, the practical requirements are those of Isaac Sim plus a GPU capable of running RTX rendering.
We recommend creating a fresh Python 3.11 environment and installing Isaac Sim before installing BeTTER itself.
Using conda:
conda create -n better python=3.11
conda activate better
Or using venv:
python3.11 -m venv .venv
source .venv/bin/activate
pip install isaacsim[all,extscache]==5.0.0 --extra-index-url https://pypi.nvidia.com
From the repository root:
pip install -e .
Additional Python dependencies used by the released tooling are collected in the project-root requirements.txt:
pip install -r requirements.txt
Use the Task Editor to define a task template instance, inspect task files, and refine object-level metadata before asset instantiation.
Use the retrieval server together with the Assets Editor to search for concrete assets, compare candidates, and update object registry content.
Use the Variation Editor to inspect and refine variation layouts, goal predicates, and fail predicates after task structure and asset instantiation are ready.
Use the task-aware teleoperation script to load a BeTTER task episode, choose a concrete variation, and record human demonstrations against the resolved task conditions.
Use the rendering script to replay recorded demonstration pickle files and export MP4 videos from the fixed BeTTER camera views.
pip install -e . is enough; install requirements.txt when you need the additional released dependencies.This release focuses on the task authoring, asset curation, variation editing, teleoperation, and rendering workflow. Additional tasks, benchmark components, and evaluation utilities will be added as they are validated.
The BeTTER source code is released under the MIT License; see LICENSE.
Some files in this repository are third-party assets or vendored third-party software and are not covered by the BeTTER MIT License. In particular, the included RoboLab-derived Franka + Robotiq asset is distributed under CC BY-NC 4.0 terms. See THIRD_PARTY_NOTICES.md for the current third-party notices and license boundaries.
If you find BeTTER useful in your research, please cite:
@article{xu2026unmasking,
title={Unmasking the Illusion of Embodied Reasoning in Vision-Language-Action Models},
author={Xu, Haiweng and Zheng, Sipeng and Luo, Hao and Zhang, Wanpeng and Xi, Ziheng and Lu, Zongqing},
journal={arXiv preprint arXiv:2604.18000},
year={2026}
}
We thank the open-source community, especially the developers of Objaverse, MimicGen, and InfiniGen, whose tools and assets have been useful for this project. In particular, our scene assets are generated with InfiniGen.
$ claude mcp add BeTTER \
-- python -m otcore.mcp_server <graph>