ForceHarness lifts frozen GPT-6 Astra from 3.3% to 90%
TL;DR
- ForceHarness raises GPT-6 Astra's zero-shot success rate on a real arm without a wrist force sensor from 3.3% to 90.0%, without fine-tuning or demonstrations.
- The harness closes the force loop below the model with self-terminating primitives, then returns each force trace as physical quantities with uncertainty.
- In simulation it adds 17.5 points on ForceBench and 34.3 points on a new seven-task MuJoCo benchmark called HiddenPhys.
On a real arm with no wrist force sensor, a frozen GPT-6 Astra's zero-shot success rate on contact-rich manipulation climbs from 3.3% to 90.0% once it runs inside ForceHarness, according to a new arXiv paper from Ke Sun, Kevin Yuchen Ma and collaborators. No fine-tuning. No demonstrations.
The paper frames the problem as two obstacles facing an API-only model. "Contact evolves much faster than an API model can decide, so the system must close the loop on force beneath the model," the authors write, and "a frozen model cannot interpret raw force streams, so force must be returned in a form the model can reason about." ForceHarness answers the first with self-terminating force primitives that regulate force on their own and return at a contact event, and the second with a physics translation layer that turns each force trace into physical quantities with uncertainty before handing it back as the tool result.
In simulation the lifts are smaller but still striking: 17.5 percentage points on the public ForceBench suite and 34.3 points on HiddenPhys, a seven-task MuJoCo benchmark the team built with hidden physical variables. The authors report the harness "surpassing force-aware policies trained on demonstrations on both benchmarks and matching hand-written expert controllers on HiddenPhys."
An ablation notes that "smaller models make significantly better decisions when given physical quantities rather than raw force values," which points at the physics translation layer as doing much of the lifting. The abstract does not say how many trials underpin the real-arm numbers, nor which tasks were run on the physical setup.
Originally reported by paper
Read the original article →Original headline: ForceHarness Lifts GPT-6 Astra Robot Success From 3% to 90% — No Fine-Tuning