On September 2, Dr. Sun Yang, the head of the Center for Intelligent Robotics X, joined StarDust Intelligence, focusing on research and application exploration in the direction of robotic reinforcement learning.
Over the past two years, the robot-based model has largely addressed whether or not silicon will fail. However, as robots move from Demo to real deployment and commercialization, the problem is shifting to whether or not silicon can stabilize. How to make robots continue to modify strategies based on real execution results, and how to return reinforcement learning to the industry's focus.
He has a long-standing experience in research and industrial practice in cross-robot reinforcement learning, large-scale distributed RL systems and large-scale model reinforcement learning. His addition will further strengthen the layout of stellar dust intelligence in the direction of robot reinforcement, and collaborate with AI models, embedded OS and rope-driven robot bodies, driving robots to continue to learn and evolve in the real world.
01.
More than a decade of reinforcement learning from tense to bytes
Dr. Sun Yi graduated from Tsinghua University and subsequently pursued postdoctoral studies at Cornell University and Rutgers University. Over the past decade, his research and industrial practice have always revolved around reinforcement learning and intelligent agents, and gradually expanded from robotic reinforcement learning to large-scale RL systems and large-scale model training, forming a technical accumulation of both algorithmic research and systems engineering.
Earlier, during his tenure at the Tencent AI Lab and Robotics X Robotics Laboratory, he was the head of the Intelligent Body Center, conducting deep reinforcement learning and robotic control research. He used deep reinforcement learning and anti-robot training to train rotary robots to achieve end-to-end proactive goals; while also researching and developing the interstellar competing AI body TStarBots, TStarBotX, and achieving important results in multi-intelligent body reinforcement MARL complex gaming environment research.
After joining the byte jump, Sun Chang expanded his technical practice from algorithmic to large-scale RL system engineering. He led the development of the core reinforcement learning infrastructure ByteRL, supporting the efficient training of multiple self-developed games AI. His team won the IEEE CoG 2023 Strategy Card AI competition.
With the rise of big-language models, he further expanded reinforcement learning accumulation to post-LLM training. During the Byte AI Lab/ByteResearch project, he was involved in the research and release of the reinforced fine-tuning method ReFT (Reinforced Fine-Tuning) and the mathematical reasoning intelligent body DeltaProver; he was deeply involved in the RLHF and pre-training of the model in the Seed team and gained deep front-end exploration experience in model alignment, reasoning enhancement and Agent direction.
From robotic reinforcement learning, to large-scale RL systems, to large-scale post-modelling training, Sun's technological path over the past decade has always revolved around a core question: how to make intelligent agents continue to learn through environmental interaction and feedback, and continuously optimize their strategies. Today, as AI moves from the digital world further into the real physical world, this set of capabilities is also finding new ground in the field of artificial intelligence.
02.
Star-dust smart coding robot trained to close the ring
In terms of stellar-dust intelligence, while advancing the construction of an AI model + OS + rope-driven core silicon technology system, Sun Yang's addition was a key component of the silicon after the certification of the robot model silicon training.
Since its inception, StarDust Intelligent has been committed to Design for AI, which believes that robotic bodies, data and AI models should not be cracked. Over the past few years, the idea has gradually formed a complete loop of training that covers core models and commercial models, front-end coexisting agents, and reinforcement learning.
The Lumo series, uniquely viewing the Lumo movement as the first civic model to continually advance robotic understanding, prediction and generation of environment and action, and Lumo-1 as the first family of stealth world action models, pioneered Latent World Dynamics, which pioneered the prediction of future worlds in stealth space and regenerated them into stealth movements; the front-end Agent Philia, which focuses on long-term memory, mission management, multi-bot interaction with nature; and reinforcement learning as a basic model, which is an important link in how robots learn from real-world continuous interaction and mission feedback and continually optimize their own strategies for real-world behavior.
The long-term accumulation of robotic reinforcement learning, large-scale RL systems and large-scale RLHF, reinforcement mitigation and Agent direction will further enhance the technical capabilities of StarDust intelligence in this regard. In the future, he will be involved in the construction of StarDust models, robotic reinforcement learning and post-training systems, exploring proven post-training methods of reinforcement learning in the Big Model era, how to further combine with real robots to perform data closures and long-term deployments.
For the past decade, I've been doing reinforcement learning and smart bodies, and I've also experienced the process of reinforcement learning from robotic controls, complex games to post-model training. Now, I really want to bring these accumulations back to the physical world.