TranscEngram Raises Funds for Embodied AI in Robotics

TranscEngram, a promising artificial intelligence startup, has successfully closed an angel funding round, securing several hundred million yuan (equivalent to billions of Japanese yen). The investment was led by prominent entities including CP Group's subsidiary SBP GROUP, alongside contributions from PDVC, Zhangjiang Hi-Tech Park, and HON-Flex. This capital infusion is earmarked for advancing the company's research and development in explainable AI models and physical world simulations. A significant portion will also support the creation of data infrastructure for multimodal motion data crucial for humanoid robots and the expansion of its talent pool.

Founded in September 2023 by AI luminaries including Professor Ma Yi, former Dean of the University of Hong Kong's Faculty of Engineering, TranscEngram is dedicated to developing foundational models for general-purpose embodied AI. The company's core mission is to build integrated systems that enhance robot perception, prediction, and action capabilities. Professor Ma posits that current large language models, while possessing vast knowledge from static data, lack the crucial ability to self-correct and validate through real-world interaction. True autonomous AI, he argues, requires a mechanism for continuous learning and internal model refinement based on accumulated experience from physical actions.

To address this, TranscEngram has architected its system with a novel dual-layer memory structure: 'visual memory' and 'motor memory.' The visual memory component functions as the robot's 'cerebrum,' interpreting spatial relationships, geometric structures, and object interactions. Complementing this, the motor memory acts as the 'cerebellum,' learning movement trajectories, time-series data, and control laws for high-frequency, stable motion execution. This sophisticated architecture aims to bridge the gap between abstract knowledge and practical physical execution, a critical hurdle in advancing robotics.

The company emphasizes its commitment to transparency with a 'white-box' network design. By extracting task-specific rules from interaction data, TranscEngram reduces reliance on manual annotation and predefined task lists. Their memory-based generative 'cerebellum' architecture has demonstrated over a threefold improvement in multi-task performance compared to traditional Vision-Language-Action (VLA) models. Internal evaluations show a success rate exceeding 95% for tasks like making coffee, tea, and folding clothes, all handled by a single, adaptable model.

A key differentiator for TranscEngram is the transferability of its learned skills across different robotic platforms. The system supports zero-shot learning, enabling robots to acquire new behaviors from a single demonstration, regardless of variations in robot hands or grippers. This adaptability is expected to significantly lower learning and deployment costs in industrial settings. The company is also developing a suite of four product lines: EngramTeleOp for low-latency remote operation, EngramEgo for collecting full-body motion data via wearable devices, EngramControl for extracting reproducible motion memories from demonstrations, and EngramNav for navigation and obstacle avoidance in unstructured environments.

Initially, TranscEngram is targeting the hotel services and high-end manufacturing sectors. In hospitality, the focus is on standardizing tasks like key card issuance and laundry services, with plans to expand to comprehensive service provision. For advanced manufacturing, particularly in aerospace, the 'cerebrum-cerebellum' system will be integrated into existing collaborative robot arms and digital production lines to enhance flexibility in high-mix, low-volume production and improve quality control. The company is actively collaborating with robotics manufacturers such as AGIBOT, Fourier Intelligence, Galbot, and DexForce to integrate its AI system into their hardware.