Former ByteDance, Tencent AI Researcher Sun Peng Joins Robotics Startup Xingchen Intelligence
On September 2, 2026, Sun Peng, a former reinforcement learning expert at ByteDance and former head of Tencent's Robotics X agent center, officially joined robotics startup Xingchen Intelligence to lead technical R&D and application exploration in robot reinforcement learning. His work will focus on post-training for robots, addressing the shift from demonstrating capabilities to ensuring stable real-world performance. Sun's expertise spans robot RL, large-scale distributed RL systems, and LLM post-training, including development of ByteRL and the ReFT method.
Coverage timeline
机器之心新闻资讯
9月2日，前字节跳动强化学习专家、前腾讯Robotics X智能体中⼼负责⼈孙鹏博⼠正式加入星尘智能，将重点负责机器人强化学习方向的技术研发与应用探索。 过去两年，机器人基础模型主要在解决“会不会”。但随着机器人从 Demo 走向真实部署和商业化，问题正在转向“能不能稳定做好”。如何让机器人根据真实执行结果持续修正策略，也让强化学习重新回到行业关注的中心。 孙鹏长期深耕强化学习（RL）、智能体（Agent）及多智能体强化学习（MARL），拥有横跨机器人强化学习、大规模分布式 RL 系统与大模型强化学习的研究与产业实践经历。他的加入将进一步强化星尘智能在机器人强化学习方向的布局，并与 AI 模型、具身 OS 和绳驱机器人本体形成协同，推动机器人在真实世界中持续学习、不断进化。 据了解，孙鹏加入星尘智能后，继续扩充机器人强化学习团队，推动相关技术能力提升和应用部署。 从腾讯到字节 深耕强化学习十余年 孙鹏博士毕业于清华大学，此后先后在康奈尔大学和罗格斯大学从事博士后研究。过去十余年，其研究与产业实践始终围绕强化学习与智能体展开，并逐步从机器人强化学习拓展至大规模 RL 系统与大模型后训练，形成了兼具算法研究与系统工程的技术积累。 早期在腾讯 AI Lab 及 Robotics X 机器人实验室期间，孙鹏曾担任智能体中心负责人，开展深度强化学习与机器人控制研究。他曾利用深度强化学习和对抗博弈训练轮式机器人，实现端到端主动目标跟随；同时研发《星际争霸》AI 智能体 TStarBots、TStarBotX，在多智能体强化学习（MARL）复杂博弈环境研究中取得重要成果。 加入字节跳动后，孙鹏将其技术实践由算法全面拓展至大规模 RL 系统工程。他主导开发了核心强化学习基础设施 ByteRL，支撑多款自研游戏 AI 的高效训练。其团队曾夺得 IEEE CoG 2023 策略卡牌 AI 竞赛冠军，研发的《炉石传说》AI 展现出击败行业顶尖选手的竞技水平。 随着大语言模型兴起，孙鹏进一步将强化学习积累拓展至LLM 后训练。在字节 AI Lab/ByteResearch 期间，他作为项目负责人参与研究并发布强化微调方法 ReFT（Reinforced Fine-Tuning）及数学推理智能体 DeltaProver；在 Seed 团队深度参与模型 RLHF 与预训练，在模型对齐、推理增强
Related stories
Astribot unveils SmoothRL for online RL during asynchronous robot execution
Astribot's foundation model team released SmoothRL, an online reinforcement learning framework designed for asynchronous execution, where robots continue acting while models compute the next action chunk. It addresses the mismatch between planned and executed actions in online RL and was validated on real high-dynamic throwing tasks, showing online learning can adapt to continuous acceleration and precise release without pauses.
1 reports · 1 sources
Alibaba, Tencent, ByteDance Restructure AI Office Agent Teams
In early July 2026, Alibaba merged QoderWork, Wukong, and MuleRun into Qianwen Office, led by DingTalk's Chen Yusen, while ByteDance integrated Feishu's product team into Doubao and its sales teams into Volcano Engine. Tencent has been consolidating resources into WorkBuddy since March, which reportedly reached an average of one million daily active users. The moves reflect a broader industry shift from coding agents to general office agents targeting white-collar workers.
2 reports · 1 sources
Google shifts AI control from DeepMind to corporate, Brin pushes Gemini and RSI
Google is restructuring its AI efforts, moving control from DeepMind back to Silicon Valley and transferring several non-technical teams from DeepMind to Google's corporate reporting structure, according to Reuters and other sources. Co-founder Sergey Brin has urged key AI staff to focus on Gemini and pushed resources toward recursive self-improvement (RSI), while Demis Hassabis has been moved to a chairman role. Reports also suggest Gemini 3.5 Pro may have been quietly canceled, with Gemini 4 already in pretraining.
ByteDance reportedly pretraining AI model with up to 10T parameters
Sources told the Financial Times that ByteDance is pretraining an AI model with up to 10 trillion parameters, roughly three times larger than Moonshot's Kimi K3 and exceeding the 8T estimate for Anthropic's Mythos 5. LatePost separately reported that ByteDance is discussing training a model with over 5 trillion parameters, led by Seed Foundation head Xiang Liang, with founder Zhang Yiming emphasizing no distillation and not being swayed by short-term trends.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe