Embodied AI Glossary中文RoboChallenge
CommonA real-robot online evaluation platform from Dexmal and Hugging Face; its first benchmark is called Table30.
RoboChallenge is a real-robot evaluation system launched in October 2025 by Dexmal (原力灵机) together with Hugging Face. Participants don't have to hand over model weights; instead, they run their model on their own machine and remotely control a real robot on the platform through an API. Its first benchmark, Table30, contains 30 tabletop manipulation tasks that probe difficulties such as precise 3D localization, occlusion, and dependence on earlier steps, using four robot types: UR5, Franka Panda, a Cobot Magic ALOHA dual-arm setup, and ARX-5. Each task is rolled out 10 times, reporting success rate and a stage-based progress score. It targets the problem that simulation scores don't equal real-robot capability, and that different labs' real-robot setups aren't otherwise comparable.
ExampleThe RoboChallenge paper used Table30 to compare π0, π0.5, CogACT, and variants of the OpenVLA series, under both single-task training and generalist training setups.
- Also called
- Table30
- Related
- Real-World Evaluation · Success Rate · Progress Score · Evaluation Protocol · Dexmal · RoboArena
- Sources
- RoboChallenge: Large-scale Real-robot Evaluation of Embodied Policies (arXiv 2510.17950)
RoboChallenge 论文 HTML 全文 (Chinese) - As of
- 2025-10