ForceHarness Takes Frozen GPT-6 Astra from 3.3% to 90.0% on a Real Robot Arm

Found first: a primary source the press has not covered yet.

Ke Sun, Kevin Yuchen Ma, Yike Zhang, Yanzhe Chen, and Mike Zheng Shou at the National University of Singapore have built ForceHarness, a three-layer adapter that gives a frozen frontier model access to force feedback during contact-rich robot manipulation, without fine-tuning or demonstration data. On a physical arm with no wrist force sensor, it lifts GPT-6 Astra's zero-shot success rate from 3.3% to 90.0%.

What the source says

ForceHarness sits between robot hardware and the model API. Self-terminating force primitives execute fast force-based control loops below the API decision rate and halt automatically on contact events. A physics translation layer converts each raw force trace into physical quantities with uncertainty estimates, then returns those as a tool-call result to the frozen model. On HiddenPhys, a seven-task MuJoCo simulation benchmark, the approach adds 34.3 percentage points, reaching 96.4%, surpassing force-aware policies trained on roughly 200 demonstrations per task. On the public ForceBench suite it adds 17.5 percentage points. Smaller models gain more from the translated physical quantities than from raw force values.

Why it matters

Frontier multimodal models handle vision-based robot tasks but consistently fail at contact-rich manipulation because force signals are inaccessible through a standard API. ForceHarness resolves this entirely on the system side, leaving frozen weights untouched and requiring no additional training data. The 90.0% result on a real arm is competitive with hand-written expert force controllers and exceeds trained force-aware policies. The gains hold across model scale, with smaller models showing larger relative improvements from physical quantity translation over raw force values.