Former Bosch Autonomous Driving Algorithm Engineer Starts Up, Using Synthetic Data to Build Haptic Large Models
Dayan Technology has secured tens of millions in angel funding to develop tactile large models and synthetic data solutions for embodied AI. The company addresses the bottleneck in robot data acquisition by combining low-cost synthetic data with proprietary "Shadow Gauntlet" haptic hardware. Their core innovation is a tactile large model that integrates physical constraints into latent space to predict force direction and optimal grasping postures. Revenue exceeded 10 million RMB in Q1 2025, dri
Analysis
TL;DR
- Dayan Technology has secured tens of millions in angel funding to develop tactile large models and synthetic data solutions for embodied AI.
- The company addresses the bottleneck in robot data acquisition by combining low-cost synthetic data with proprietary "Shadow Gauntlet" haptic hardware.
- Their core innovation is a tactile large model that integrates physical constraints into latent space to predict force direction and optimal grasping postures.
- Revenue exceeded 10 million RMB in Q1 2025, driven by cost-effective data collection from real-world scenarios like unmanned supermarkets.
- The strategy leverages a closed-loop ecosystem where custom hardware feeds high-quality data into proprietary models, creating significant technical barriers.
Why It Matters
This development highlights the critical shift in embodied AI from visual-centric perception to multi-modal sensory integration, specifically addressing the "data hunger" of robot brains. By demonstrating that synthetic data combined with specialized haptic feedback can significantly reduce training costs and improve precision in complex tasks like grasping, it offers a scalable pathway for commercializing humanoid robots. This approach challenges the industry's reliance on expensive, low-efficiency real-world data collection methods.
Technical Details
- Tactile Large Model: A multimodal model that incorporates physical constraints within its latent space to output precise force vectors and optimal grasping poses, filling a gap in domestic public offerings.
- Proprietary Hardware ("Shadow Gauntlet"): A self-developed haptic glove featuring 29 array units and 1,015 tactile contact points with a 300Hz response frequency, capable of simultaneously capturing hand posture and force feedback.
- Synthetic Data Pipeline: Utilizes 4D world reconstruction and controllable diffusion models to generate training data at a fraction of the cost of manual labeling (cents per frame vs. yuan), achieving over 60% gross margins.
- "Heterogeneous Purchase" Data Collection: Partners with unmanned supermarkets and front warehouses to collect first-person perspective data from human workers wearing their devices, mapping human actions to robotic controls with high efficiency.
- Closed-Loop Architecture: Deep coupling between hardware data formats, calibration processes, and model training algorithms ensures that the value of the hardware is intrinsically linked to the proprietary software ecosystem.
Industry Insight
- Data Efficiency Over Volume: The success of synthetic data and human-to-robot mapping suggests that future AI development will prioritize data quality and generation efficiency over sheer volume of raw real-world captures, drastically lowering entry barriers for robotics startups.
- Hardware-Software Synergy as Moat: The integration of specialized sensors with tailored AI models creates a defensible competitive advantage; competitors copying hardware alone cannot replicate the performance without the corresponding algorithmic ecosystem.
- Expansion into Global Markets: With plans to establish subsidiaries in markets like Saudi Arabia, companies focusing on foundational AI infrastructure (data and models) are well-positioned for international expansion, leveraging local government support and global demand for automation.
Disclaimer: The above content is generated by AI and is for reference only.