Paper review · November 19, 2026
ForceHarness: A harness that lets frozen frontier models use force information as actions and physical quantities
ForceHarness: Turning Force into Action and Physics for Frozen Frontier Models
The authors built a harness with force calibration, self-terminating force primitives and a physical-quantity translation layer, enabling API models to perform contact-rich robot manipulation without training.
Problem
Frontier multimodal models can control robots zero-shot from images alone. But for tasks where contact force matters, such as tightening screws or comparing weights, images are not enough. Existing force-aware policies require demonstration data and weight updates, so they cannot be used with frozen API-only models. Contact also changes faster than the model can make decisions, and models cannot interpret raw force signals.
Method
It has three layers. (1) A calibration layer that estimates the force on the hand from a wrist sensor or joint torques (Jacobian least squares) and re-zeros at rest. (2) 12 force primitives (contact search, push, slide, spiral search, insertion, screw tightening, lifting, etc.) that maintain force with admittance control on top of a base impedance controller and stop on their own when a contact event occurs. (3) A translation layer that converts force trajectories into physical quantities such as mass, friction coefficient and contact height, along with uncertainty and status flags, and returns them as tool results. For evaluation, the authors also built HiddenPhys, a MuJoCo benchmark of 7 tasks with hidden physical variables.
Results
- On 6 tasks with a real Franka FR3 without a wrist force sensor, GPT-6 Astra's zero-shot success rate rose from 3.3% without force information to 90.0%
- On the public ForceBench (RGB input), it rose from 70.0% to 87.5%, surpassing force-aware policies trained on that benchmark
- On HiddenPhys, it rose from 62.1% to 96.4%, comparable to a hand-crafted expert force controller and well ahead of Diffusion Policy, π0.5 and ForceVLA trained on about 200 demonstrations per task
- In ablations, the authors reported that smaller models made much better decisions when given physical quantities rather than raw force values
Limitations
The real-robot evaluation covers 6 tasks on a single robot (Franka FR3). In HiddenPhys, object poses were given as noisy text to separate out visual perception. Instructions were written by people for each task, and the 12 primitives were also designed by people. Because the model makes a decision once every few seconds, there may be limits for dynamic tasks where speed matters.
Why it matters
The paper shows a practical design for giving API models a "sense of touch" without collecting demonstrations or fine-tuning. The finding that frozen models do better with physically interpreted values than with raw sensor readings is also worth considering when adding other sensors.
Key terms
- Harness
- A software layer that wraps a frozen model and provides tools, perception and control. It extends capabilities without changing the model weights.
- Impedance and admittance control
- Impedance control produces force by treating the gap between the target and actual positions like a spring. Admittance control converts force error into a target velocity. Combining the two keeps force at the desired value.
- Self-terminating primitive
- A unit of motion that stops on its own when an event such as contact or completed insertion occurs, rather than after a set time. It handles fast contact control in place of the slow model.
Original
Authors: Ke Sun, Kevin Yuchen Ma, Yike Zhang, Yanzhe Chen, Mike Zheng Shou · Posted 2026-11-16 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.