Skip to content
Bud's Wiki

Open-X-Tactile — community effort to build the world's largest heterogeneous tactile manipulation dataset

TL;DR

Section titled “TL;DR”

Michael Yuan (Tsinghua IIIS, working with Yang Gao) announces Open-X-Tactile, a community-organized initiative to assemble the largest heterogeneous tactile manipulation dataset under a unified data format and benchmark ecosystem, explicitly positioned as the tactile analog of Open X-Embodiment. The tweet reports 50+ institutions and companies already involved — including UC Berkeley, MIT, CMU, Tsinghua, ETH Zürich, NUS, and the tactile-hardware vendors Sharpa, Wuji, Paxini, and Xense. Contributions are accepted at any scale, across any tactile sensor / robot embodiment / manipulation setup. The initiative is being organized privately with a preview website, and Yuan is soliciting contributors via email or reply.

Key claims

Section titled “Key claims”
  • The goal is a heterogeneous tactile manipulation dataset that spans multiple sensor modalities, robot embodiments, and manipulation setups under one data format [tweet body].
  • 50+ institutions and companies are already involved, spanning academia (UC Berkeley, MIT, CMU, Tsinghua, ETH Zürich, NUS) and the tactile-hardware vendor ecosystem (Sharpa, Wuji, Paxini, Xense) [tweet body].
  • Contributions of any dataset size, tactile sensor, robot embodiment, or manipulation setup are considered valuable — no minimum bar [tweet body].
  • Framing: “the tactile equivalent of Open X-Embodiment” — i.e. the community-scale cross-embodiment aggregation move applied to tactile signals [tweet body].
  • The initiative is currently organized privately with a preview site linked from the tweet; entry is via email (ycb24@mails.tsinghua.edu.cn) or a reply [tweet body].

Method

Section titled “Method”

Tweet-level announcement, not a paper. The methodological content is the coordination mechanism: a unified data format spec plus a benchmark ecosystem into which heterogeneous contributors can drop tactile-sensor + robot + task episodes without pre-committing to a single sensor type, taxel layout, or embodiment. No public spec of the format has been shared at filing time. Public engagement (replies from PaXini Tech, SharpaRobotics, community members) confirms the initiative is being read as a serious cross-vendor coordination move rather than a single-lab data push.

Results

Section titled “Results”

None yet — this is an announcement of an in-progress data collection, not a released dataset or benchmark. The tweet reports 6.2K views and public confirmation from at least two of the named vendors (PaXini, Sharpa) within a day of posting.

Why it’s interesting

Section titled “Why it’s interesting”

Every filed instance under Tactile sensing for manipulation has flagged the large-scale tactile dataset gap: Data Pyramid for Embodied Manipulation names it as open challenge #1 of six for embodied-manipulation data recipes, Sharpa releases the Origami Dataset for the IROS 2026 Robotic Origami Challenge is a single-vendor / single-task instance filling part of that gap (682 episodes on paper folding), and Being-H0.8: A Latent Tactile World-Action Model at Scale resorts to inferring tactile from egocentric human video for lack of ground-truth-tactile at pretraining scale. Open-X-Tactile is the first filed attempt to solve this at the community-coordination layer rather than by a single lab. It also sidesteps the sensor-format-fragmentation question that Tacmap: Bridging the Tactile Sim-to-Real Gap via Geometry-Consistent Penetration Depth Map and Tactile Genesis: Exploring Tactile Sensors at Scale for Learning Dexterous Tasks disagree on (Tacmap commits to one abstraction — deform-map penetration depth; Tactile Genesis unifies 8 abstractions under one simulator) by punting sensor-heterogeneity into the format itself. Worth watching whether the released format spec picks a “one abstraction” or “many abstractions under a superset schema” bet — that decision downstream-conditions which of the filed simulator / policy / vision-only-substitute recipes can consume it.

See also

Section titled “See also”