PHANES AI
Category: Robotics / Embodied AI
Chinese embodied-AI startup building tactile-aware world models for humanoid whole-body mobile dexterous manipulation, trained on human vision, touch and motion data. PHANES AI was founded in 2026. The company is led by Shuo Yang (杨朔). Based in Shanghai, China. Team size: 11-50. Latest round: Angel. Key investors include 雅瑞资本 (Yarui Capital).
- Founded
- 2026
- Headquarters
- Shanghai, China
- Team size
- 11-50
Value proposition
Raises the learning value per hour of human demonstration data by capturing synchronized egocentric vision, hand tactile pressure and full-body motion, then converting it into robot-ready multimodal training data for world models and manipulation policies.
Products and solutions
SeePhys Hub, a whole-body full-modality wearable data capture system launched with Tinno Mobile in October 2026: 4-camera headband (360° view), two wrist bands with cameras, three waist and leg motion trackers, and tactile gloves, plus a proprietary data-processing and motion-retargeting platform. Research line: EgoTouch (egocentric visuo-tactile dataset), TouchAnything (tactile estimation from video), TouchWorld (predictive and reactive tactile foundation model) and HumanWBC (whole-body loco-manipulation control). A full-modality world model is planned for 2027.
Unique value
Combines whole-body motion and dense hand touch with egocentric vision in one portable rig, filling the gap between cheap internet/Ego video and expensive motion-capture or teleoperation data, backed by a tactile world model validated on real robots.
Target customer
Humanoid and dexterous-robot makers (robot body OEMs), embodied-AI model developers, and scenario partners in industrial, food-service and retail settings.
Industries served
Robotics / embodied AI, humanoid robots, industrial manufacturing, food service and retail
Technology advantage
Full-stack loop of hardware, data pipeline, tactile world model and real-robot evaluation. Headband ≤220g, wristbands ≤75g, 24h continuous capture, in-headband sync error ≤0.5ms. TouchWorld reached 65.0% average success on 6 real-robot tasks (53.7% with perturbations), beating the strongest baseline among Pi-0.5, FTP-1 and GR00T N1.7 by 15.7 and 18.5 points.
How they differentiate
Focuses on touch plus whole-body motion rather than vision-only Ego or UMI collection; uses a hierarchical tactile world model (1Hz planning / 10Hz action / 30Hz tactile correction) instead of folding touch into a single VLA; does not build its own robot body, partnering with robot OEMs instead.
Main competitors
大晓机器人 (environment-based data collection + world model), TARS 它石智航 (human-centric embodied data), Generalist AI (large-scale human data pretraining)
Key partnerships
Tinno Mobile (天珑移动), co-launched SeePhys Hub, Harbin Institute of Technology (Shenzhen)
Major milestones
2026-05: company founded. 2026-07: TouchWorld tactile foundation model released. 2026: Shenzhen subsidiary 破晓智能（深圳）有限公司 established (RMB 15M registered capital). 2026-10-14: SeePhys Hub launched with Tinno Mobile. 2026-10: angel round from 雅瑞资本. 2027 (planned): full-modality world model.
Growth metrics
Team of several dozen people; over 90% of core members from Tsinghua, Peking University, HKU and HIT. EgoTouch dataset covers 208 tasks, 1,891 interactions, 20+ hours of video, ~2.1M frames and 1,000+ objects.
Market positioning
Early-stage, research-led data and world-model layer for humanoid dexterous manipulation, targeting the middle tier of a three-layer data pyramid between internet video and high-precision mocap or real-robot data.
Geographic focus
China (Shanghai, Shenzhen)
Patents and IP
Research papers: TouchWorld (arXiv 2607.07287) and TouchAnything (arXiv 2605.13083). Team plans to open-source EgoTouch, TouchAnything and TouchWorld data, code and models.
About Shuo Yang (杨朔)
Tenured Professor and PhD supervisor, School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen); named tenured professor at age 26 (born 1998). Google PhD Fellowship recipient; PhD work was an ICLR Best Paper Finalist; national-level young talent program. Research covers multimodal LLMs, data-centric AI, trustworthy AI, CV and embodied AI.
Latest news about PHANES AI
More Robotics / Embodied AI companies
Official website: https://phanes-lab.github.io/TouchWorld-website/