The tactile layer of world models

Tactile World Model Tracker

Name: Tactile World Model Tracker
Creator: RoboSkin.ai
License: https://creativecommons.org/licenses/by/4.0/

Explore robot world models that predict or use tactile, contact and force information. Compare what each model predicts, the hardware it uses, and the evidence behind its real-robot evaluation. Records are updated as primary sources are checked.

Review preview — editorial approval pending

These source-backed records are available for review. Assessments and evidence grades are provisional, and authors have not confirmed them. This page is excluded from search indexing until editorial review is complete.

Models tracked: 8
Latest source check: 2026-09-30

Explore the records →

Compare model evidence

Newest first by initial paper date. Grades describe available evidence, not performance. “Not verified” means unknown, not that an artifact does not exist.

Predicted modalityEvidence gradeCode availability

Showing 8 of 8 models

Source-backed tactile world model records. Expand a model to review its sources and limitations.
Model / sources	First published	Predicts	Tactile sensor	Robot platform	Real-robot evaluation	Open artifacts	Evidence
InternW0 Sources and assessment Provisional assessment: Asynchronous world prediction supports contact-aware post-training; laboratory progress scores do not establish full-workflow success or isolate tactile contributions. InternW0: A Foundational Physical World Model for Efficient Real-World Interactions Institutions: Shanghai AI Laboratory Author check: Not contacted. Sources checked: 2026-09-30. Source 1: arxiv.org/abs/2609.27656v1 Source 2: arxiv.org/html/2609.27656v1 Source 3: internrobotics.github.io/internw0 Source 4: roboskin.ai/news/internw0-contact-aware-world-model-pipetting Read the RoboSkin brief / guide →	First published2026-09-23	PredictsVision, Touch / contact, Action	Tactile sensorNot verified	Robot platformDual-arm parallel-gripper platform; 7-DoF arm with 20-DoF hand for pipetting	Real-robot evaluation Author-reported real-robot evaluation Tasks: 5 Trials per setting: 15	Open artifacts Code: Not verified Data: Not verified Weights: Not verified License: Not verified	D Provisional
Agile-WAM Sources and assessment Provisional assessment: Asymmetric visual and tactile prediction supports five physical tasks; source-reported latency is model timing, and code is announced rather than released. Agile-WAM: An Agile Tactile World Action Model for Contact-Rich Robot Control Institutions: University of California, Davis; Analog Devices Author check: Not contacted. Sources checked: 2026-09-30. Source 1: arxiv.org/abs/2609.20761v2 Source 2: arxiv.org/html/2609.20761v2 Source 3: hanchuzhou.github.io/TARO_project_page/ Source 4: roboskin.ai/news/agile-wam-tactile-world-action-model-robot-control Read the RoboSkin brief / guide →	First published2026-09-17	PredictsVision, Touch / contact, Action, Latent representation	Tactile sensorAnalog Devices 32 x 32 piezoresistive pressure sensor	Robot platformFlexiv Rizon 4	Real-robot evaluation Author-reported real-robot evaluation Tasks: 5 Trials per setting: 20	Open artifacts Code: Announced Data: Not verified Weights: Not verified License: Not verified	B Provisional
DexTouch-WM Sources and assessment Provisional assessment: Human pretraining improves reported contact prediction, but generated trajectories give mixed real-robot policy scores; code availability remains unverified. DexTouch-WM: Learning Action-Conditioned Tactile World Models from Human Touch for Dexterous Robot Manipulation Institutions: HKUST (GZ); Xspark AI; PKU; THU; HKU Author check: Not contacted. Sources checked: 2026-09-30. Source 1: arxiv.org/abs/2609.20649v2 Source 2: arxiv.org/html/2609.20649v2 Source 3: roboskin.ai/research/dextouch-wm-human-touch-world-model-2026 Read the RoboSkin brief / guide →	First published2026-09-17	PredictsVision, Touch / contact	Tactile sensorMoxian piezoresistive tactile gloves; shared 320-taxel training layout	Robot platformTianji arm with 20-DoF Wuji hand	Real-robot evaluation Author-reported real-robot evaluation Tasks: 4 Trials per setting: 10	Open artifacts Code: Not verified Data: Not verified Weights: Not verified License: Not verified	D Provisional
HiTac-WAM Sources and assessment Provisional assessment: Hierarchical contact forecasts support candidate selection and corrective replanning on three real-robot tasks; implementation availability remains unverified. HiTac-WAM: A Hierarchical Tactile World Action Model for Contact-Rich Robot Manipulation Institutions: Institute of Automation, Chinese Academy of Sciences; ImprintX Robotics; Beijing Academy of Artificial Intelligence Author check: Not contacted. Sources checked: 2026-09-30. Source 1: arxiv.org/abs/2608.19574v1 Source 2: arxiv.org/html/2608.19574v1 Source 3: roboskin.ai/research/hitac-wam-hierarchical-tactile-world-action-model-2026 Read the RoboSkin brief / guide →	First published2026-08-20	PredictsTouch / contact	Tactile sensorBilateral DM-Tac W2 sensors	Robot platformIMETA-Y1	Real-robot evaluation Author-reported real-robot evaluation Tasks: 3 Trials per setting: 30	Open artifacts Code: Not verified Data: Not verified Weights: Not verified License: Not verified	D Provisional
FeelWorld Sources and assessment Provisional assessment: Contact-aware planning is evaluated on three real-robot tasks; code availability remains unverified, so the availability-based evidence grade is provisional. FeelWorld: Visuo-Tactile World Model for Hierarchical Contact Prediction and Planning Institutions: Institute of Automation, Chinese Academy of Sciences; Imprintx Robotics; Beijing Academy of Artificial Intelligence Author check: Not contacted. Sources checked: 2026-09-30. Source 1: arxiv.org/abs/2607.24267v1 Source 2: arxiv.org/html/2607.24267v1 Source 3: roboskin.ai/research/feelworld-visuo-tactile-world-model-2026 Read the RoboSkin brief / guide →	First published2026-07-27	PredictsVision, Touch / contact, Latent representation	Tactile sensorDM tactile sensors producing 3D tactile point clouds	Robot platformImeta-Y1	Real-robot evaluation Author-reported real-robot evaluation Tasks: 3 Trials per setting: 40	Open artifacts Code: Not verified Data: Not verified Weights: Not verified License: Not verified	D Provisional
ViTacWorld Sources and assessment Provisional assessment: The v2 study evaluates generated-data policy improvement over 50 trials per policy, stage and task; official code remains announced. ViTacWorld: Scaling Visuo-Tactile World Models for Contact-Rich Robot Manipulation Institutions: ShanghaiTech University; InstAdapt Author check: Not contacted. Sources checked: 2026-09-30. Source 1: arxiv.org/abs/2607.22530v2 Source 2: arxiv.org/html/2607.22530v2 Source 3: vitacworld.github.io/ Source 4: roboskin.ai/guides/visuo-tactile-world-models-robot-manipulation Read the RoboSkin brief / guide →	First published2026-07-24	PredictsVision, Touch / contact	Tactile sensorXense optical fingertip sensors	Robot platformFranka Panda with Robotiq 2F-85	Real-robot evaluation Author-reported real-robot evaluation Tasks: 4 Trials per setting: 50	Open artifacts Code: Announced Data: Not verified Weights: Not verified License: Not verified	B Provisional
TouchWorld Sources and assessment Provisional assessment: Predictive subgoals and reactive correction are evaluated on six tasks; per-condition rollout counts and implementation availability remain unresolved. TouchWorld: A Predictive and Reactive Tactile Foundation Model for Dexterous Manipulation Institutions: Harbin Institute of Technology, Shenzhen; PHANES AI Author check: Not contacted. Sources checked: 2026-09-30. Source 1: arxiv.org/abs/2607.07287v2 Source 2: arxiv.org/html/2607.07287v2 Source 3: phanes-lab.github.io/TouchWorld-website/ Source 4: roboskin.ai/news/touchworld-tactile-foundation-model-dexterous-manipulation-2026 Read the RoboSkin brief / guide →	First published2026-07-08	PredictsVision, Touch / contact	Tactile sensorJQ-Industries tactile glove	Robot platformUnnamed humanoid with Wuji dexterous hands	Real-robot evaluation Author-reported real-robot evaluation Tasks: 6 Trials per setting: Not disclosed	Open artifacts Code: Not verified Data: Not verified Weights: Not verified License: Not verified	D Provisional
Dream-Tac Sources and assessment Provisional assessment: Joint action and observation generation has six-task real-robot evidence and public code; cross-sensor transfer and independent replication remain unestablished. Dream-Tac: A Unified Tactile World Action Model for Contact-Rich Robot Manipulation Institutions: Peking University; The Hong Kong University of Science and Technology; Nanjing University Author check: Not contacted. Sources checked: 2026-09-30. Source 1: arxiv.org/abs/2606.08737v1 Source 2: arxiv.org/html/2606.08737v1 Source 3: github.com/LYFCLOUDFAN/Dream-Tac Source 4: roboskin.ai/research/dream-tac-tactile-world-action-model-2026 Read the RoboSkin brief / guide →	First published2026-06-07	PredictsVision, Touch / contact, Action, Latent representation	Tactile sensorTwo Xense Photon fingertip sensors	Robot platformFranka Emika Panda	Real-robot evaluation Author-reported real-robot evaluation Tasks: 6 Trials per setting: 20	Open artifacts Code: Released Data: Not verified Weights: Not verified License: Apache-2.0 (code only)	A Provisional

How to read the evidence

A — Real robot and released code: Author-run physical evaluation plus a verified public implementation. This does not imply independent replication.
B — Real robot, code explicitly unreleased: Physical evaluation with an explicit coming-soon or no-release statement.
C — Simulation only: The reviewed evaluation is confined to simulation.
D — Insufficient evidence for a complete grade: This includes real-robot papers whose code release status is unknown. Read the separate evaluation column; D does not mean no robot experiment was reported.

Trials are counted per policy, task and evaluation condition. TouchWorld reports 100 rollouts per task without a condition split, so its per-setting count stays unknown. ViTacWorld uses the v2 50-trial protocol.

Scope, sources and corrections

Inclusion requires a versioned primary source connecting robotic world prediction to touch, contact or force. General video generation, games and cinematic world building are outside scope. Predicted modalities are outputs rather than conditioning inputs; HiTac-WAM is classified by its tactile forecast head.

Different robots, tasks and protocols are not a shared leaderboard. Source review does not establish reproducibility, deployment safety or transfer to another sensor. Missing artifact links stay unknown; article licenses do not license models or datasets.

RoboSkin.ai’s original tracker compilation is offered under CC BY 4.0. Cite “RoboSkin.ai Tactile World Model Tracker” with this page URL and your access date. Third-party papers, code, weights and data retain their own terms.

Submit a model or correction →