EVENT DOSSIER
To address the issue of integrating semantic action generation with physical interaction control in contact-rich mobile manipulation tasks, researchers proposed the FWBC-VLA framework. This framework introduces a sensorless residual torque estimator (HSR-Force) to infer contact intensity and its temporal changes, and encodes contact information as tokens to be injected into the VLA action expert to perceive contact states. In mobile-manipulation tasks, the pre-trained VLA backbone network is fine-tuned using the WL&Arm dataset containing over 5,000 scenarios, and the robot’s proprioception, Jacobian-derived force estimates, and contact state are jointly inputted into a compensation generator to generate corrective actions. Finally, the combined operation center actions and corrective actions are transmitted to the WBC policy execution. The framework was verified as effective in real-world experiments such as whiteboard wiping and opening doors with door closers.
arXiv:2609.03889v2 FWBC-VLA 提出力感知框架用于接触丰富操作
提出 FWBC-VLA 框架，旨在解决接触丰富型移动操作任务中语义动作生成与物理交互控制之间的衔接问题。该框架通过引入无传感器残差扭矩估计器 HSR-Force 推断接触强度及其时间变化，并将接触信息编码为 token 注入 VLA 动作专家以感知接触状态。在移动 - 操作任务中，对预训练 VLA 骨干网络使用包含超过 5000 个场景的 WL&Arm 数据集进行微调，并将机器人本体感觉、雅可比导出的力估计及接触状态联合输入补偿生成器产生校正动作。最终将操作中心动作与校正动作结合后传递给 WBC 策略执行。在白板擦拭和带闭门器的门开启等真实世界实验中验证了该框架的有效性。