About the project

ForceBench: Benchmarking Force Integration for Robotic Manipulation

Xuan Han^*, Zinuo Jin^*, Tianlin Li^*, Yunfan Lou^*, Xin Liu, Yuchen Xue, Yuanpeng He, Boxian Yao, Shuhan Zhan, Ruihai Wu, Zhibo Pang

^* Equal contribution.

A simulation benchmark for evaluating six-axis wrist-force integration under a unified protocol, with a complementary real-world study.

What ForceBench studies

Force feedback can guide action selection and execution, but differences in tasks, demonstrations, and control settings make comparisons between integration designs difficult. ForceBench brings these comparisons into a shared setting.

The benchmark combines eight tasks, 1,600 successful demonstrations, and task-specific success criteria across physical-property inference, resistance-dependent interaction, and contact-rich assembly. The submitted study compares one no-force baseline and six force integration designs on a shared π_0.5 backbone, alongside additional reference systems.

A shared experimental setting

Task-specific training, common demonstration data, matched evaluation seeds, and explicit observation and action interfaces support comparisons across methods.

Beyond a single mean

Outcome diagnostics distinguish physical decisions from complete execution. The real-world study examines whether the main task-dependent trends also appear on a physical robot.

Reading the results

The site reports the submitted manuscript’s simulation and real-world results. Prompt Fusion has the highest observed mean simulation success, while Prompt Fusion and Dual-Path Fusion share the highest observed mean in the real-world study.

These results show task-dependent benefits and limitations. Temporal systems are separate reference designs, rather than a controlled history-length ablation. Alternative backbones and systems also retain differences described in the protocol.

Results and comparison scope Experimental protocol

Citation

The manuscript is not an accepted conference publication. The entry below identifies the current submitted work.

@misc{han2026forcebench,
  title = {ForceBench: Benchmarking Force Integration for Robotic Manipulation},
  author = {Han, Xuan and Jin, Zinuo and Li, Tianlin and Lou, Yunfan and
            Liu, Xin and Xue, Yuchen and He, Yuanpeng and Yao, Boxian and
            Zhan, Shuhan and Wu, Ruihai and Pang, Zhibo},
  year = {2026}
}

Project resources

Release availability is listed on the project homepage.

The task overview and experimental protocol describe the submitted study.