ForceHarness lifts GPT-6 Astra from 3.3% to 90% on force-based robot tasks

Researchers at the National University of Singapore (NUS) have presented ForceHarness, a system that lets frozen frontier models — used only via API, with no further training — control robots on tasks that depend on the sense of touch. On a real robot arm, OpenAI's GPT-6 Astra jumped from 3.3% to 90.0% success on contact-rich manipulation tasks, with no demonstrations and no fine-tuning. The work was published as a preprint on arXiv on November 16, 2026.

The problem ForceHarness solves is well known: multimodal models already control robots from images, but many tasks — fitting a part, pushing an object with the right force, feeling when something touches — depend on contact forces that images cannot show. And there are two obstacles to bringing force to an API-accessed model: contact changes much faster than the model can decide, and a frozen model cannot interpret a raw stream of force data.

How ForceHarness works

The system has three layers. First, twelve self-terminating force primitives: small controllers that regulate force on their own, in a fast loop below the model's decision rate, and return when a contact event happens. Then, a physics translation layer converts each force trace into physical quantities — such as mass, friction and contact height — with uncertainty estimates. Finally, those quantities return to the model as a tool-call result, in a form it can reason about.

Hand force can be calibrated from a wrist force sensor or from joint torques alone — in the real-arm experiment, there was no wrist sensor. The authors also built HiddenPhys, a seven-task benchmark in the MuJoCo simulator with hidden physical variables, to compare the system with expert controllers and trained policies.

The results

According to the authors, in simulation ForceHarness added 17.5 percentage points on ForceBench, a public benchmark, and 34.3 points on HiddenPhys, surpassing force-aware policies trained on about 200 demonstrations per task and matching hand-written expert controllers on HiddenPhys. On the real arm — a Franka FR3 with no wrist force sensor — GPT-6 Astra's zero-shot success rate on six contact-rich manipulation tasks rose from 3.3% to 90.0%.

The tests also indicate that smaller models make much better decisions when given physical quantities instead of raw force values — a detail that could make this kind of control cheaper. The numbers come from the authors themselves, in the preprint, and the abstract does not state how many trials support the real-arm result. The paper has not yet been peer-reviewed, so the results should be read as a promising demonstration, not as an established fact.

Why it matters

The result arrives as OpenAI dissolved its robotics team to bet on general models such as GPT-6 Astra commanding robots. ForceHarness is a concrete example of how that path can work: instead of training a model just for robots, the system gives the API model what it lacked — the sense of force — without touching the model's weights.

If the results hold up in independent evaluations, the approach could speed up general-purpose robots controlled by frontier models, without needing to collect hundreds of demonstrations for each new task. The system's code was released as open source, so other labs can reproduce and test the results.

Image credit: Franka Emika robot arm at a trade fair, photo by Ims, CC BY-SA 4.0, via Wikimedia Commons (file: Franka Emika2.jpg). The photo is illustrative: the paper's tests used a Franka FR3 arm.

Sources

Transparency: this article was produced with the help of artificial intelligence, cross-checked against public posts on X and internet sources. The facts were verified in the sources listed above; please consult the original sources to confirm details.

By Muse