Embodied In-Context Learning: A Magnet for Entrepreneurs, Context as the New Pathway for Scaling
48 minute ago / Read about 0 minute
Author:小编   

In 2020, the In-Context Learning (ICL) technology of GPT-3 sent ripples through the Natural Language Processing (NLP) landscape. Today, this innovative approach is quietly making its mark in the realm of embodied intelligence. In August of this year, Skild AI unveiled its robotic foundation model, S1, which is capable of mastering new tasks without the need for fine-tuning, simply by observing a task demonstration video once. Similarly, Generalist AI's GEN-1.5 model boasts One-Shot Learning capabilities.

Domestic startup Keke Matrix is at the forefront of establishing ICL as the foundational paradigm for embodied intelligence. Its founder, Gao Yuxiang, brings a wealth of experience from his previous tenure at Fourier, where he played a pivotal role in developing full-sized humanoid robot systems (Note: The original text's "connect through" has been interpreted and integrated into a more natural English phrase). In contrast to earlier methods in embodied intelligence that heavily depended on vast datasets, the current industry trend places a premium on robots' learning and adaptive capabilities, positioning ICL as a promising new avenue for development.

However, the application of ICL in the embodied domain still grapples with three significant technical hurdles: visual comprehension, multimodal fusion, and the processing of long-context and memory. To overcome these obstacles, Keke Matrix has ingeniously positioned ICL during the pre-training phase, dynamically extracting information through the construction of 'conditional representations' and adhering to the principle of 'strong comprehension, light generation.' Presently, the company has achieved a robust One-Shot adaptation completion rate exceeding 80% for straightforward tasks, while also delving into the exploration of robots' Self-Correction abilities.

Looking ahead, Keke Matrix will concentrate on reducing the time required for robots to acquire new tasks. It's worth noting that the company's name, Keke Matrix, is a creative amalgamation of four English words: COEXIST, COWORK, COLLABORATE, and COEVOLVE, reflecting its vision and ethos.