Wrench-ACT makes force and torque the robot policy action
Wrench-ACT pairs force-reflecting bilateral demonstrations with an ACT policy that directly commands a six-dimensional target wrench.

Researchers from Siemens and the University of Technology Nuremberg released Wrench-ACT on September 29, 2026. The system modifies the usual imitation-learning contract: instead of predicting a target position and letting controller error create contact force, its Action Chunking with Transformers policy outputs a six-dimensional force/torque wrench directly. With matching force-reflecting demonstrations, it averages 76.8% success across five contact-rich tasks versus 40% for a position-policy baseline. Paper and version record.
Explore tactile datasets for the broader comparison.
Key takeaways
- Direct wrench output works best only when the demonstrations contain intentional wrench commands. Converting ordinary position demonstrations into approximate wrenches averages 28.8% in the same action space.
- The main table evaluates four collection/action combinations on five tasks with 50 rollouts each: 1,000 physical evaluation rollouts in total.
- The paper promises more than 1,000 wrench-action demonstrations on a companion site upon publication, but no public companion URL, archive or license was verified with the v1 release.
What changed
Position policies command where the tool should go. During contact, a Cartesian impedance controller turns the pose error into force. That is useful but indirect: two identical target poses can produce very different interaction loads as geometry and stiffness change.
Wrench-ACT removes the pose target from the learned action. Three RGB streams, measured follower wrench, robot position and velocity, and gripper state enter a single-task ACT model. Its seven-dimensional output contains a six-axis target wrench plus the gripper command. The inner robot controller runs at 500 Hz, policy inference near 50 Hz, and observations are collected at 60 Hz. The force-mode interface retains velocity damping; the learned wrench is not a direct motor-torque command. The latest policy target is held between predictions. Method and controller.
Data collection is the other half of the design. In the bilateral setup, a human pushes a leader arm and feels the follower's reaction. The measured leader wrench becomes the action label. A second dataset uses a Meta Quest position interface and Cartesian impedance control. The authors also translate each dataset into the opposite action representation, producing four policy conditions that separate the collection interface from the learned output.
Results and the matched-data effect
The five tasks are peg insertion, fuse clipping, fan insertion, industrial-connector mating and pen writing. In the peg setup, a person holds the peg to ensure grasp acquisition; the transfer and insertion phases are autonomous. The peg result therefore does not measure autonomous pickup-to-insertion. Bilateral data plus wrench output records 74%, 88%, 58%, 80% and 84%, averaging 76.8%. The conventional VR-data/position-action condition records 6%, 96%, 8%, 12% and 78%, averaging 40%.
That headline average hides two qualifications. Wrench-ACT is lower on fuse clipping, 88% versus 96%, and only three task differences are statistically significant after the paper's multiple-comparison correction: peg, fan and industrial-connector insertion. Pen writing and fuse clipping do not establish a significant advantage.
The cross-condition table is more revealing. A wrench policy trained on wrenches reconstructed from VR position trajectories averages 28.8%. A position policy derived from bilateral wrench data averages 34.8%. The strongest outcome appears when the interface and action agree: humans deliberately command force, and the policy predicts the same quantity.
An inference-rate ablation on the industrial connector reports 80% at 50 Hz, 36% at 30 Hz, 20% at 15 and 5 Hz, and 12% at 1 Hz. The nominal result uses 50 trials; each reduced-rate condition uses 25. When the paper support under the pen is raised 2.5 centimeters, the bilateral-wrench policy falls from 84% to 76%, while the VR-position policy falls from 78% to 24%.
What this means for robotics
RoboSkin analysis: Wrench-ACT is evidence for aligning a teleoperation interface with the quantity a policy must control. Merely logging a force/torque sensor alongside position commands does not mean the dataset contains deliberate force strategy. For contact-rich data, action provenance matters as much as the presence of force channels.
Direct wrench control is not a universal replacement for pose actions. The authors deliberately choose tasks where interaction force is central, and they state that incidental-contact tasks may not benefit. The system also depends on a capable inner force loop and a mechanically compliant setup; the learned policy is only one layer of the control stack.
Limitations and availability
Wrench-ACT is an arXiv v1 preprint, and RoboSkin.ai has not reproduced it. All experiments use one UR5e-based setup and single-task models trained from scratch. Datasets contain 200–300 episodes per task and interface, but operator count is not stated. The force-reflecting collection rig is substantially more specialized than common handheld or VR capture systems.
The v1 introduction and abstract promise demonstrations upon publication, while its conclusion describes a release alongside the work. The November 4 arXiv record check did not resolve a concrete companion archive or its terms. Public payload, code, checkpoints and reuse licenses remain unverified; neither sentence alone proves access.
Record assistance and outcome alongside force
Low force can mean that the task was never completed. The paper summarizes a high-percentile force from each rollout, and discusses low-force fan failures that did not finish insertion. That statistic is not a damage test or a general safety result. Evaluation details.
For a comparison on your task, keep three records together: who established the initial grasp; whether the required insertion, release or writing outcome occurred; and the force trace with units, reference frame and timestamps. Declare any reset or intervention before counting a trial. Keep the measured wrench, demonstrated target wrench and controller-adjusted command in separate fields.
This proposed review prevents an easier starting condition or an incomplete low-force attempt from looking like improved manipulation. It does not add tests to the authors' study. Use the control-interface comparison before choosing a demonstration format.
Sources and related resources
- Wrench-ACT arXiv v1, submitted September 29, 2026
- Full controller, tasks and cross-condition results
- RoboSkin guide to robot manipulation
- RoboSkin Physical AI and touch guide
*Evidence review updated November 4, 2026. The original paper and article dates are retained; source inspection is not a RoboSkin.ai reproduction.*


