| --- |
| license: apache-2.0 |
| task_categories: |
| - robotics |
| tags: |
| - vla |
| - vision-language-action |
| - benchmark |
| - semantic-grounding |
| --- |
| |
| # RoboSemanticBench (RSB) |
|
|
| [RoboSemanticBench (RSB)](https://huggingface.co/papers/2606.02277) is an embodied benchmark for diagnosing **semantic grounding in action prediction** for vision-language-action (VLA) models. |
|
|
| In each episode, a robot receives a multiple-choice math or general-knowledge question, observes candidate answer blocks, and must grasp the block corresponding to the correct answer. RSB is designed to separate low-level grasping ability from the model's capacity to use instruction semantics to select the correct physical target. |
|
|
| ## Links |
|
|
| - **Paper:** [RoboSemanticBench: Diagnosing Semantic Grounding in Action Prediction for VLA Models](https://huggingface.co/papers/2606.02277) |
| - **Repository:** [https://github.com/ZGC-EmbodyAI/RoboSemanticBench](https://github.com/ZGC-EmbodyAI/RoboSemanticBench) |
| - **Project Page:** [https://robotwin-platform.github.io/](https://robotwin-platform.github.io/) |
|
|
| ## Benchmark Suites |
|
|
| RSB covers six embodied semantic-answering suites: |
|
|
| | Suite | Task name | Choices | Semantic source | |
| | --- | --- | ---: | --- | |
| | RSB-Math-4 | `rsb_math` | 4 | Procedural arithmetic | |
| | RSB-Math-10 | `rsb_math_10blocks` | 10 | Procedural arithmetic | |
| | RSB-HardMath-4 | `rsb_hardmath` | 4 | GSM8K | |
| | RSB-HardMath-10 | `rsb_hardmath_10blocks` | 10 | GSM8K | |
| | RSB-General-4 | `rsb_general` | 4 | MMLU-style QA | |
| | RSB-General-10 | `rsb_general_10blocks` | 10 | MMLU-style QA | |
|
|
| ## Metrics |
|
|
| The benchmark reports three diagnostic metrics: |
| - **Task Success Rate (TSR)**: The episode succeeds only when the robot selects the correct block and places it in the answer zone. |
| - **Grasp Success Rate (GSR)**: The robot successfully grasps any candidate block, regardless of correctness. |
| - **Normalized Semantic Grounding (nSG)**: Semantic target selection conditioned on successful grasping, revealing the gap between motor competence and semantic grounding. |
|
|
| ## Sample Usage |
|
|
| ### Data Collection |
| To collect expert trajectories (e.g., for `rsb_math`): |
| ```bash |
| bash collect_data.sh rsb_math rsb_math_train_500 0 --skip_instructions |
| ``` |
|
|
| ### Policy Evaluation |
| To evaluate a policy using the RoboTwin interface: |
| ```bash |
| python script/eval_policy.py --config policy/Your_Policy/deploy_policy.yml \ |
| --overrides \ |
| --task_name rsb_general_10blocks \ |
| --task_config rsb_general_10blocks_test_500 \ |
| --ckpt_setting your_checkpoint_name \ |
| --seed 0 \ |
| --policy_name Your_Policy |
| ``` |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{RoboSemanticBench, |
| title={RoboSemanticBench: Diagnosing Semantic Grounding in Action Prediction for VLA Models}, |
| author={Bin Yu and Yao Zhang and Haishan Liu and Shijie Lian and Yuliang Wei and Xiaopeng Lin and Zhaolong Shen and Changti Wu and Ruina Hu and Bailing Wang and Cong Huang and Kai Chen}, |
| year={2026}, |
| eprint={2606.02277}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.RO}, |
| url={https://arxiv.org/abs/2606.02277}, |
| } |
| ``` |
