预约贾鹏GTC2026讲面向灵巧操作的高效强化学习框架直播

Core Insights - The company ZhiJian DongLi has completed five rounds of financing in less than six months, raising a total of 2 billion RMB [1] - The CEO of ZhiJian DongLi is Jia Peng, who previously led the autonomous driving technology development at Li Auto [2] - The Chairman is Wang Kai, former CTO of Li Auto, and the COO is Wang Jiajia, who was responsible for mass production of intelligent driving at Li Auto [3] Company Strategy and Technology - The current embodied foundational models struggle to achieve near 100% success rates while adhering to strict execution timelines in real-world tasks [4] - Reinforcement learning is viewed as a key technological pathway to bridge this gap, yet it faces core challenges such as sparse rewards, low sample efficiency, and high real-world trial-and-error costs in dexterous operation scenarios [4] - A new efficient reinforcement learning framework has been proposed, built on a unified embodied foundational model that integrates visual and language understanding, image generation, action generation, and value generation within the same architecture [5] - This unified model paradigm enhances cross-temporal and cross-modal joint reasoning capabilities, significantly improving the generalization ability of strategies [5] - The framework provides process-level dense information for reinforcement learning, which includes not only actions but also visual and linguistic latent spaces, thereby improving the efficiency of reinforcement learning in real-world dexterous operations [5] - The learning framework has been deployed on NVIDIA Thor, demonstrating its potential for scaling and mass application of embodied foundational models [5]