ForceHarness Gives Frozen Multimodal Models Force Feedback
The ForceHarness paper introduces a system that lets frozen API-based frontier multimodal models use force feedback for robotic manipulation without demonstrations or fine-tuning. It combines self-terminating force primitives with a physics translation layer that converts force traces into physical quantities with uncertainty. The paper reports that on a real arm without a wrist force sensor, GPT-6 Astra’s zero-shot success rate rose from 3.3% to 90.0%, while simulation gains were 17.5 percentage points on ForceBench and 34.3 points on HiddenPhys.
The reported results suggest that force feedback can substantially expand what frozen multimodal models can handle in contact-rich manipulation without task-specific demonstrations or fine-tuning.
Computer Science > Robotics
arXiv:2611.21099 (cs)
[Submitted on 16 Nov 2026]
Title:ForceHarness: Turning Force into Action and Physics for Frozen Frontier Models
Authors:Ke Sun, Kevin Yuchen Ma, Yike Zhang, Yanzhe Chen, Mike Zheng Shou
View a PDF of the paper titled ForceHarness: Turning Force into Action and Physics for Frozen Frontier Models, by Ke Sun and 4 other authors