Skip to content

Source paper

ForceHarness: Turning Force into Action and Physics for Frozen Frontier Models

Ke Sun, Kevin Yuchen Ma, Yike Zhang and 2 more

Ke Sun, Kevin Yuchen Ma, Yike Zhang, Yanzhe Chen, Mike Zheng Shou

arXiv 2611.21099 · 202640 upvotes on Hugging Face

Research ideas from this paper (3)

Generated by AI. The paper’s authors did not write these.

AI summary

ForceHarness enables API-only, frozen frontier multimodal models to perform contact-rich robotic manipulation without demonstrations or fine-tuning. It introduces a three-layer architecture comprising force sensing and calibration, self-terminating closed-loop force primitives, and a physics translation layer that converts force traces into interpretable physical quantities. Across real-robot hardware and simulation benchmarks including HiddenPhys and ForceBench, the system dramatically improves zero-shot manipulation success and matches expert force controllers.

How it works (AI summary)

The framework places three modular layers between a frozen API model and low-level robot control: a bottom sensing layer estimating hand forces from wrist sensors or external joint torques via Jacobian pseudo-inversion; a middle execution layer running admittance control loops at 20 Hz that regulate force targets and terminate on discrete physical events (contact, stall, resistance drops); and a top translation layer computing physical estimates (mass, friction, stiffness, contact height) with standard errors to pass back as tool results.