First-hand information for everyone
A team of researchers has released a preprint describing Spark-0, a new approach that aims to make robots better at contact-rich, dexterous tasks by incorporating large-scale tactile data into pretraining. The work — reported in arXiv:2610.22342v1 and accompanied by a dataset and benchmark — suggests that touch can serve two roles: as supervision that improves visual representations, and as a live input that helps controllers react to force and slip.
Why touch? Modern generalist robot policies have benefited from large-scale visual pretraining, but they still struggle when tasks depend on forces and very local contact dynamics that cameras cannot reliably show. The Spark-0 team addressed this by collecting HumanTouch, a human-manipulation dataset with about 1,000 hours of bilateral full‑hand tactile recordings synchronized with egocentric video and hand pose, captured from 32 operators across more than 100 tasks. The same glove design can be mounted on the target robot hand, which the authors say helps align human and robot signals.
Two complementary uses of touch are central to Spark-0. During pretraining, the researchers convert glove signals into ‘‘contact maps’’ that are rendered into the camera view and used together with a frozen 3D-geometry teacher to supervise intermediate visual features — a train-only signal that is removed at deployment. Separately, a tactile encoder summarizes recent regional forces and taxel (tactile pixel) states and feeds those tokens into the action-generation pathway, so the policy can be deployed with live touch if available. The authors emphasize that a single trained checkpoint supports both vision-only deployment (when the robot has no tactile sensors) and tactile-conditioned deployment.
To evaluate the approach, the team also released SparkArena: a paired evaluation suite that includes vision-only simulated tasks and contact-rich real-robot tasks on the same dexterous embodiment. They compared Spark-0 to a range of prior vision-language-action and tactile policies on SparkArena and on the Ego Humanoid Manipulation Benchmark. According to the preprint, Spark-0 achieved the highest mean success rates on both SparkArena tracks and led on all aggregate metrics of the Ego Humanoid benchmark.
How big is the reported effect? The paper highlights two headline numbers comparing vision-only pretraining (same RGB data but with tactile signals removed) to tactile-aware pretraining: retaining touch from the full 1,000-hour HumanTouch dataset increased vision-only success on the Ego Humanoid benchmark from about 52.5% to 71.7%. On the real tactile deployment track of SparkArena, success reportedly rose from 44.0% to 52.0% when touch was used at deployment and retained during pretraining. These improvements are evidence that tactile data can act as scalable supervision to strengthen visuomotor representations, not only as an extra sensor at runtime.
Why be cautious? The results are promising but should be viewed in context. The findings come from a preprint rather than a peer‑reviewed publication, and independent replication will be important. The improvements appear strongest when large volumes of touch data (the full 1,000 hours) are used; smaller amounts of tactile pretraining produced weaker or inconsistent gains in some settings. The SparkArena real‑robot evaluation uses a particular dexterous hand and a glove layout that matches the robot; how well the approach generalizes to other hands, sensors, or environments remains to be shown.
There are practical and safety considerations as well. Tactile sensors and full‑hand gloves add hardware cost and complexity; real-world deployment would require robust, high‑bandwidth tactile sensing and careful handling of sensor faults, occlusions, and calibration differences between human and robot data. The paper’s contact-map supervision relies on reconstructing hand geometry from glove data — an extra modeling step that can introduce error if the reconstruction or registration is imperfect. Finally, the authors note that some pretraining objectives are removed at inference, which simplifies deployment but means the benefit depends on having representative pretraining data.
Still, the work points to an attractive idea for advancing robot manipulation: using human tactile demonstrations at scale to teach vision-based models where and how contact happens. If the results hold up across more hardware and in independent tests, the combination of visual pretraining with tactile supervision could improve the robustness of dexterous manipulation in settings such as assembly, tool use, or household chores.
The authors have made datasets and code available (project page and repo cited in the preprint). Next steps for the community will include reproducing the gains on other robots and sensors, testing long‑term robustness, and exploring how tactile pretraining interacts with safety constraints and real‑world uncertainties.