GLoTouch lets one parallel gripper search and identify objects without external vision

Zonglin Li and colleagues introduced GLoTouch on September 23, 2026, as an arXiv v1 preprint from an independent researcher, Shanghai Jiao Tong University and the University of Hong Kong. The system gives a standard parallel gripper two haptic roles: it first carries a passive probe to search a container with wrist force sensing, then puts the probe away and uses bilateral visuotactile fingertips to identify and retrieve the object matching a supplied 3D mesh. In 50 randomized real-robot trials, the complete pipeline retrieved 38 targets. Paper and version record.

Explore tactile datasets for the broader comparison.

Key takeaways

• GLoTouch retrieved 84 of 100 targets in simulation and 38 of 50 on a Flexiv robot. Each scene contained the same five 3D-printed object classes, and success required search, recognition, grasping and lifting.
• The local matcher needs no object-specific training, but it does need the target STL model, known container registration, known object count and objects that are reachable from above.
• The paper says code will be open-sourced. No repository, downloadable dataset or software license was linked from the v1 record on September 24. Full methods and experiments.

What changed

Parallel grippers normally trade dexterity for simple, reliable grasping. GLoTouch extends their perceptual range without adding an actuated search hand. During global exploration, the Xense gripper holds a spherical-tipped probe that has no embedded sensor. A six-axis wrist force-torque sensor, robot kinematics and the known probe shape localize contact while the robot follows a coverage path. Multi-directional probing estimates candidate centers, contours and heights.

The robot then places the probe in a holder. Bilateral visuotactile depth maps and jaw aperture form a local TouchSet for each candidate. A finite-window matcher compares width, depth and acquisition mode against synthetic observations generated from the target STL. The decision checks absolute matching quality, the best-versus-second-best margin and sufficient tactile evidence before grasping; it does not simply choose the lowest loss regardless of evidence.

This division matters: the probe covers a larger workspace, while the fingertips collect spatially resolved geometry only after an object is found. The same gripper remains available for the final grasp.

Results under the reported conditions

In simulation, each of five targets was tested 20 times in randomized five-object layouts. GLoTouch completed 84 of 100 retrievals. A separate global-search ablation found all five objects within a 5-meter motion budget in 55% of 100 layouts, compared with 30% for random descent. That is a scene-discovery metric, not the end-to-end success rate.

The local matching component was also isolated on 60 target-and-phase samples from five physical objects. Width and action alone achieved 41 of 60 correct top-1 matches. Adding local depth raised this to 47 of 60; the complete matcher reached 54 of 60. These are development-set samples that reuse different subsets of contact sequences, so 60 should not be read as 60 independently collected objects or an unseen-object test.

On hardware, every target was evaluated in 10 randomized scenes. The Flexiv arm and Xense gripper retrieved 38 of 50 targets, or 76%. Apple was retrieved in all 10 trials, while Train succeeded in 5 of 10. Confusion between Train and Truck, object motion during probing and toppling during clamping were reported failure modes.

What this means for robotics

RoboSkin analysis: GLoTouch is a useful integration pattern for haptic search. It reserves high-resolution visuotactile sensing for the stage where local geometry is valuable, and uses cheaper tool-mediated force contact for global coverage. A detachable probe may be easier to service than a large tactile array and leaves the end effector free for manipulation.

The method is not touch-only in the broadest sense: it relies on proprioception, wrist wrench sensing, known robot-to-container registration and a pre-existing target mesh. It also assumes top access and objects resting on the container floor. Deployments in cluttered bins would need safeguards for entanglement, stacked objects and accumulated pose error. Those integration conditions are at least as important as the recognition score for teams building tactile manipulation systems.

Limitations and availability

This is a preprint, not an independently reproduced result, and RoboSkin.ai has not run the system. Evaluation uses one arm-gripper configuration and five printed objects. The global stage depends on a wrist force-torque sensor; the passive probe does not turn an otherwise sensorless gripper into a contact detector. The paper also assumes a known scene cardinality and does not evaluate transparent versus opaque objects as separate conditions because external vision is excluded from control.

The November 4 review of the arXiv record and v1 paper still did not verify a linked implementation repository, data archive or applicable release license. The manuscript’s planned code release is not sufficient evidence of an available, reusable payload. This is a limit of the reviewed sources, not a claim that no other release can exist.

Define when the contact evidence is insufficient

A low matching loss alone is not a trustworthy object identity. GLoTouch combines tactile depth, aperture and acquisition geometry, then checks evidence coverage and separation from the next candidate. The paper’s alternative losses use different scales, so their numerical margins are not interchangeable confidence values. Nor does a 54/60 development-set matcher imply that search, identification and lifting each succeed 90% of the time. Matching protocol and ablations.

For an integration review, keep three decisions separate:

• Contact localization: retain probe geometry, wrist tare, coordinate registration and whether probing moved the object. The paper’s 0.37 mm center error on one square calibration case does not establish general object-pose accuracy.
• Identity selection: retain the candidate meshes, sufficient-contact check, absolute fit and relative margin. An unfamiliar object or poor observation requires an explicit unresolved state, rather than an assumed match.
• Task completion: record the selected target, grasp, lift and any toppling or intervention. Revalidate the contact map if the object moves between search and grasp.

These are proposed receiving checks. The 84/100 simulation result includes runtime failures, while the 55% scene-discovery result measures finding all five objects within a motion budget. Neither can replace the physical 38/50 end-to-end result. Compare this known-geometry search with the other signal and decision roles in the manipulation guide, rather than treating “without external vision” as an absence of cameras, geometry priors or calibration.

Sources

• arXiv v1 record, submitted September 23, 2026
• Full GLoTouch v1 paper

*Source review updated November 4, 2026. Original publication dates are retained. Results belong to the cited authors; RoboSkin has not run these systems or downloaded their experimental payloads.*