Daimon‑Infinity Tactile Dataset Open‑sourced

Robots usually learn from cameras and motor logs; Daimon Robotics has now open-sourced a touch-heavy dataset meant to teach machines what contact feels like. (autonews.gasgoo.com) Daimon said on April 16 that the first public release of Daimon-Infinity includes 10,000 hours of data and is available on Alibaba’s ModelScope platform. The company said the full project is being built with dozens of academic and industry partners and is planned to grow to millions of hours this year. (autonews.gasgoo.com) The dataset pairs tactile readings with vision, language and action, so a model can link what a robot sees, what it is told, what it does and what it feels at the same moment. Daimon describes itself as a Shenzhen company focused on high-resolution multimodal tactile sensing systems and dexterous robot hands. (modelscope.csdn.net) (dmrobot.com) Touch data matters most in contact-heavy jobs like inserting parts, gripping soft objects or adjusting force after first contact. Camera-only systems can miss slippage, pressure and small alignment errors that show up only when a gripper actually touches something. (arxiv.org) (github.com) That is the gap robotics groups have been trying to close as they build vision-language-action systems, which are models that turn images and text instructions into motor commands. Google DeepMind says its Gemini Robotics models are designed to bring multimodal reasoning into the physical world, while recent tactile-robotics papers argue that large touch datasets are still scarce. (deepmind.google) (github.com) Google DeepMind has already pushed large shared robot datasets through Open X-Embodiment, a project that standardized more than 1 million real robot trajectories from 34 labs across 22 robot types. Daimon-Infinity appears to target a narrower bottleneck: high-resolution touch, synchronized with the rest of the robot’s sensory stream. (deepmind.google) (robotics-transformer-x.github.io) Daimon and partner writeups say the new release is aimed at cross-embodiment learning, meaning a policy trained on one robot can transfer more easily to another with different hardware. The company also said the open portion is only the first batch, with a larger expansion planned by the end of 2026. (autonews.gasgoo.com) (www.c114pro.com) The harder question is whether the dataset’s claimed efficiency gains hold up outside Daimon’s own benchmarks. As of April 18, Google DeepMind’s public research pages list broad robotics work and publications, but I could not verify a separate official DeepMind paper page for Daimon-Infinity itself. (deepmind.google 1) (deepmind.google 2) For robotics researchers, the release puts touch data into the same scaling race that already reshaped vision and language. The next test is simple: whether outside labs can download the data, train on it and reproduce better manipulation on real machines. (modelscope.cn) (github.com)