Forcebench compares force feedback methods for robot grasping and assembly
ForceBench: Benchmarking Force Integration for Robotic Manipulation
Robotics
Summary
Robots can use a sense like touch, called force feedback, to better handle objects. The authors created ForceBench, a set of robot tasks and tests to fairly compare different ways robots use this sense. They found that one method called Prompt Fusion works best overall, improving how well robots complete tasks like figuring out object properties and assembling parts. They also tested these methods on real robots, confirming similar results. This work helps builders know which force feedback techniques work best for different robot skills.
What this means in practice
- •For robotic system developers: Evaluate and choose force feedback integration methods to improve robot grasping and assembly performance under diverse task demands.
- •For industrial automation teams: Test different force-based control designs to optimize robotic assembly processes, helping robots adapt to contact-rich environments.
Authors
Xuan Han, Zinuo Jin, Tianlin Li, Yunfan Lou, Xin Liu, Yuchen Xue, Yuanpeng He, Boxian Yao, Shuhan Zhan, Ruihai Wu, Zhibo Pang
Abstract
Force feedback provides vision-language-action (VLA) policies with physical observations that can guide action selection and execution. However, differences in tasks, demonstrations, and control settings across existing systems hinder reliable comparisons of force integration designs. We introduce ForceBench, a simulation benchmark for evaluating six-axis wrist-force integration under a unified protocol. ForceBench combines eight tasks spanning physical-property inference, resistance-dependent interaction, and contact-rich assembly with 1,600 synchronized demonstrations and task-specific success criteria. We compare one no-force baseline and six force integration designs on a shared $π_{0.5}$ backbone. Prompt Fusion achieves the highest observed mean simulation success of 35.31%, compared with 23.25% for the baseline, and improves success rates across all three task categories. Predictive and temporal gains in simulation concentrate on physical-property inference, while both temporal designs underperform the baseline on assembly. Real-world evaluation of eight systems across five tasks supports the main qualitative trends: Prompt Fusion and Dual-Path Fusion achieve 63.00% mean success, compared with 40.00% for the baseline, but only Prompt Fusion improves every task. These findings support ForceBench's practical value for identifying the benefits and limitations of force integration across manipulation demands. The project page is available at https://forcebench-platform.github.io/.