Spark-0: Scaling Spatially Aligned Tactile Pretraining for Robot Dexterous Manipulation
Chinese start-up Xspark AI and university co-authors recorded 1,000 hours of people handling objects while wearing tactile gloves, and used that touch data to pretrain a robot manipulation model. In this preprint the touch-pretrained model raised success on an unseen humanoid benchmark from 52.5% to 71.7% even when the robot had no touch sensors at test time, and beat the strongest tactile baseline on real contact-rich tasks by 16 points.
Can touch recorded from human hands at scale be used not only as a live input to a robot policy but also as a training signal that makes a vision-only policy better at dexterous manipulation?
Vision-language-action models are pretrained on images and video, but force, compliance and the first moment of slip are mostly invisible to cameras, which is why contact-rich jobs such as inserting test tubes or packing shoes remain hard. Robot tactile datasets are small and tied to specific hardware.
Prior vision-tactile-language-action work fed touch into the policy as an extra input, trained on relatively small robot datasets, and required tactile sensors at deployment. Generalist models such as pi0.5 and GR00T N1 rely on vision and language pretraining without touch.
The team built HumanTouch: 1,000 hours of full-hand tactile manipulation from 32 operators across more than 100 tasks, recorded with gloves carrying 320 taxels per hand and 25 finger joints at 60 Hz, mixed with egocentric video, robot and simulation data into a 1.9k-hour pretraining set. Spark-0 builds on a Qwen3-VL-4B backbone. During pretraining only, contact maps projected into the camera image and features from a frozen 3D geometry teacher supervise the visual tokens; a separate tactile encoder feeds force histories and taxel states to the action expert. One checkpoint can therefore run with or without touch.
On the authors' SparkArena benchmark, built on their Gen 1 Prototype robot (Marvin M6S arms with WUJI Hand 2 hands and Moxian tactile gloves), Spark-0 led FastWAM by 8.6 points on seven vision-only simulated tasks and the tactile policy FTP-1 by 16.0 points on five real contact-rich tasks (52.0% against 36.0% success). On the external Ego Humanoid benchmark it ranked first on all eight aggregate metrics, 5.1 points ahead of EgoVLA on short tasks and 25.1 points ahead of VITRA on long ones. Against the same data with touch discarded, keeping 1,000 hours of touch lifted Ego Humanoid success from 52.45% to 71.72% and real-robot success from 44.0% to 52.0%; a 100-hour subset did not beat the vision-only variant on two of three benchmarks.
This is a preprint without a limitations section. SparkArena is the authors' own benchmark on their own embodiment, and the real track has only five tasks, so the 16-point margin needs outside replication. The gain appears only at the full 1,000-hour scale, which implies an expensive data operation, and tactile deployment needs gloves with the same taxel layout as the human data. Simulated gains were small (+0.29 points from touch).
The paper strengthens the business case for glove-based human data collection, a field where Chinese and US start-ups and hand makers such as WUJI and Sharpa are investing, and it suggests tactile data has value even for robots that ship without skin sensors. For humanoid makers targeting assembly and packing, the practical horizon is model releases within 12 to 24 months; for Korean hand makers such as Tesollo and Wonik Robotics, matching human and robot taxel layouts becomes a design consideration.