No wrist sensor needed: ForceHarness lifts Astra from 3.3% to 90% on a real arm

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

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

cs.RO

2026-11-16

Self-terminating primitives close the force loop under a frozen API model and return physical quantities, lifting GPT-6 Astra from 3.3% to 90% on a sensorless arm.

What problem this solves

Frontier multimodal models can already drive a robot arm from images with no task training. Driving a screw, seating a USB plug, and sorting look-alike bottles by weight still depend on contact force, which the camera never shows. An API-only model with frozen weights hits two gaps. Contact changes faster than one model call, so the force loop cannot wait. The model has also never learned to read a raw wrench, the six-axis force and torque.

Force-aware policies and vision-language-action models take wrist force as a network input, so every task needs demonstrations and a weight update. Harnesses that do expose contact usually pass raw sensor numbers back, or ask the model to write the low-level force code itself.

Method

ForceHarness, from Show Lab at the National University of Singapore, places three layers between the frozen model and the robot. No demonstrations and no finetuning. Each turn the model makes one tool call. Inside the call, a primitive regulates force at controller rate and stops on a contact event. The tool result then comes back as physical quantities.

The bottom layer estimates the force the environment applies to the hand. A wrist sensor reads it directly. With no sensor, external joint torques go through the flange Jacobian by least squares. The wrench is de-biased and sign-flipped. The trace is quiet, but the zero drifts with pose, and filtering does not remove that slow bias. Before an empty-hand primitive the arm tares at rest, treating the still reading as zero. After the call it settles, averages a short window, and reports only the change within the episode. The real robot is a Franka FR3 with no wrist force sensor. Its stock impedance controller is unchanged: the pose target leads the hand like a spring, more lead means more force, and a cap on the lead caps the force.

The middle layer is twelve self-terminating force primitives, plus about ten tools for moving, grasping, and locating. Five groups cover making contact, holding force, finding and seating a part, turning under load, and measuring. Holding a force is admittance control. Force error integrates into a target velocity, and the inner spring produces the force. A primitive stops on contact, stall, a resistance drop, or a force cap, not on a timer. The model picks direction and how hard to press. The primitive picks when to return. A fixed guard limits force and speed. The search primitive presses at about 1 N along a spiral. When the reaction falls it drives downward, and reports how far the part advanced and whether it dropped in.

The top layer does not judge the task. Averages and least-squares fits turn each trace into mass, friction, contact height, stiffness, and insertion advance, with uncertainty. Mass comes from vertical force during a still hold. Friction is tangential force over normal force on the sliding portion. Whether that is enough is left to the model.

Results

Six real tasks, ten episodes each, scored by an operator from the final state. Through the Codex CLI at medium reasoning effort, GPT-6 Astra moves from 2/60 (3.3%) with no force to 54/60 (90.0%) with ForceHarness. Every task differs at p < 0.01 by Fisher's exact test. GPT-6.1 Sol, under the same prompts, reaches 47/60.

TaskAstra, no forceAstra + ForceHarnessSol + ForceHarness
Wipe a whiteboard1/108/106/10
Drive a screw0/109/107/10
Peel an eggplant1/109/109/10
Insert a USB plug0/108/107/10
Sort bottles by weight0/1010/108/10
Tell soft from hard0/1010/1010/10

RGB-D shows neither mass nor stiffness, so both property tasks are 0/10 without force. Absolute grams from joint torques drift, but order within an episode is stable, and Astra scores 10/10 on both. On the four contact tasks, wrist depth is only good to about 1 to 2 cm, and position-only motion succeeds in 2 of 40 episodes. Ink left after an 8 N wipe is retried at 10 N. Rising side load during a screw triggers an in-place release. A peel that has advanced 51 mm at 2.77 N mean force keeps its 6 N target.

On the eight ForceBench tasks, ten seeds each, Astra goes from 70.0% to 87.5%. Sol also reaches 87.5%. Claude Fable 5.2 reaches 75.0%. The strongest trained result cited there is 35.31% for π0.5 with Prompt Fusion, from about 200 demonstrations. The gain that separates statistically is Drawer Pull, 0/10 to 10/10.

HiddenPhys is a new seven-task MuJoCo benchmark. Each episode hides one of mass, center of mass, table height, hole position, latch angle, or unlock force, and code checks the final state. Astra and Fable with force both score 135/140 with zero handle damage. A hand-written expert scores 136/140 with one damage event. Without force, Astra is at 87/140 (17 damage events) and Fable at 96/140 (11). Diffusion Policy, π0.5, and ForceVLA, about 200 demonstrations each, land between 65/140 and 83/140, with 13 to 29 damage events. Adding wrist force to those policies does not help consistently. After the hidden ranges shift and only the frozen models' instructions change, Astra and Fable drop about 8% and 5% relative on four tasks (80 episodes), to 69/80 and 71/80, beside the expert at 68/80. Astra's cost per episode falls by 45%.

Ablations are simulation only. Event words without force values lift smaller GPT-5.x models from 0.18 to 0.54 and Sol from 0.63 to 0.77, and cut Lock-handle damage for the smaller models from 22 of 25 episodes to 3. Judging a held load has no stop event. Sol scores 0.87 to 0.91 on a force plot, raw numbers, or physical quantities. Smaller models reach 0.56 on plots. Raw numbers move decision accuracy from 0.57 to 0.69. Quantities with uncertainty reach 0.85 (McNemar p = 0.023 against raw numbers) and 0.70 success. One-shot programs rise from 23% to 77%. Asking the model after every primitive reaches 91%.

Why it matters

Contact is fast and the API call is slow. Leave regulation inside the primitive, and leave mass, friction, and advance in the tool result, and a frozen model can run these tasks on an arm with no wrist sensor and no new demonstrations.

On HiddenPhys the expert comparison is a tie, 135 against 136. The small-model ablation is the cleaner result. Raw samples barely move the decision. Physical quantities do.

Limitations

The paper states three limits. Seconds elapse between primitives, so force cannot be corrected continuously while a motion is underway. The menu is fixed, so a new motion law has to be written in. Ablations ran only in simulation.

The ForceBench comparison is loose. The no-force condition already scores 70%, and the instructions include facts demonstrations would have shown a trained policy. Tools and perception thresholds were tuned on eight of the ten seeds. Stove Press and Nut Thread were revised after the first runs. The public checker skips release and clearance on the faucet, stove, and drawer tasks. Trained numbers come from the original paper, 200 evaluation episodes against 10 here. The other tasks do not separate at that sample size.

HiddenPhys gives the frozen model noisy pose text. The learned policies see cameras, and the Peg hole is visible in the image. Peg is still the short task: Astra 15/20, expert 18/20. The real-robot study is ten episodes per task, labeled by an operator, with procedural hints in every instruction except soft versus hard. Joint torques support within-episode order, not absolute grams. The one-shot program comparison tuned only this harness. A 90% success rate means the split works on these six tasks.

Terms

Source

Related papers

All paper explainers