A robot trained in simulation used touch signals to insert a peg through a tight, angled fixture on a real arm, according to a new preprint. In 30 hardware trials, its full tactile controller seated the peg 19 times, compared with 14 successes when the controller used the touch signal as one combined input, rather than separate information about where contact occurred, how force spread and how contact changed over time. 
The fixture made the task sensitive to small alignment errors: a square peg had to pass through a one-millimeter-clearance bore tilted 10 degrees, then clear two internal protrusions. The team varied pickup and fixture positions. It trained policies in Isaac Sim and transferred them to a UR5e arm with a Robotiq two-finger gripper and two capacitive tactile pads, without further real-world training.
The Oct. 7 arXiv paper by Siqi Shang and colleagues addresses a mismatch between simulated contact and a physical sensor. Rather than model every pad deformation, it maps contact force and contact area to a shared “effective contact response,” then uses per-taxel calibration, which corrects each individual tactile-sensor element for its sensitivity, to recover that response from sensor readings. From this calibrated response, the controller derives three separate views: contact geometry (where touch occurs), force distribution (how load spreads across the pads), and temporal change (how the contact pattern changes over time). A mask lets the same policy use different combinations of these signals; all eight combinations are trained together, so the chosen combination can change at deployment without retraining.
The simulation results did not predict the best hardware configuration. A controller using the unfactorized response scored 70.1% in simulation, above the full three-part version at 61.3%. On the robot, the ranking reversed: the full version succeeded in 19 of 30 insertion trials, versus 14 of 30 for the single response and 5 of 30 when the controller used only the robot’s own state sensing, without tactile input (proprioception alone). That is five additional successes in this test, not evidence of a general 35% improvement across robots.
In two other tasks, geometry-focused signals helped align a held object, while force signals helped regulate grip. Together, the experiments suggest that a robot may benefit from choosing which touch cues matter for a task rather than compressing them into one summary.
The result remains an early demonstration. It is a preprint, and the physical insertion comparison covers 30 trials on one arm, gripper and sensor configuration. The authors have not established that the gains carry over to other hardware or factory conditions.
Sources
- Shang et al., “Factorized Tactile Representation and Control for Sim-to-Real Manipulation” (arXiv, Oct. 7, 2026)
- Image source and license: Wikimedia Commons (CC BY 4.0)
.webp){kind=link}
Image credit: A-Lab robotic installations for sample transfer, by the study authors, via Wikimedia Commons, CC BY 4.0. The photograph is illustrative and does not show the peg-insertion experiment.