Jie Chen (陈杰)
I am currently the Founding Director and Principal Researcher of the Large-Model Reinforcement Learning Department and the Embodied Software Systems Department at XPENG Robotics. We focus on cutting-edge technologies of embodied AI and the mass production of full-size humanoid robots, including but not limited to:
- real-time embodied agents for long-horizon tasks
- scaling for humanoid whole-body control
- humanoid loco-manipulation
- reinforcement learning training suites
- full-stack embodied evaluation and data infrastructure
- embodied simulation
Previously, I was a Senior Researcher and Tech Lead at Seed, ByteDance, focusing on post-training of Large Language Models (LLMs) and agentic RL. We made some contributions to ByteDance's Doubao and Seed series of models.
We are hiring — LLM Researchers, Embodied AI Researchers, LLM Infrastructure Experts, Reinforcement Learning Researchers, Software Architects, and Systems Engineers. Feel free to reach out.
We work on foundation models and full-stack infrastructures that move humanoids into the world.
News
| Aug 2026 | Our robotics business closed a Series A of over US$900M (valuation over US$6.3B), fueling the mass production of Physical AI. |
|---|---|
| Aug 2026 | We released DeepInsight II, tracing one trace from benchmark to robot. |
| Aug 2026 | We released Capek 0.5, an execution-centric vision-language model for embodied intelligence. |
| Aug 2026 | We released R2S-EGO, dual-proxy refinement for sparse-capture real-to-sim. |
| Jul 2026 | We released Athena-Brain, an efficient 8B robot brain for general intelligence and embodied interaction. |
| Jul 2026 | We released Athena-WBC, capability-aligned policy experts for long-tail humanoid whole-body control. |
| Jun 2026 | We released DeepInsight, a unified evaluation infrastructure across the Physical AI stack. |
| May 2026 | We released the Engagement Process to formalize the real-time embodied agent paradigm. |
| May 2026 | We released Vigil, disentangling world completion from self-termination in embodied agents. |
| Nov 2025 | Our next-generation IRON humanoid robot debuted at 2025 XPENG AI Day, targeting mass production by the end of 2026. |
| May 2025 | Seed-Coder, an open-source family of lightweight code LLMs. |
| Apr 2025 | Seed-Thinking-v1.5, a reasoning model. |
| Jan 2025 | Doubao 1.5 / 1.5-pro, an MoE model with an integrated training–inference design, rivaling GPT-4o on aggregate benchmarks. |
| Dec 2024 | verl, an open-source RL training framework for LLMs, was released. |
| Oct 2024 | SandboxFusion, an open-source secure sandbox for running and judging LLM-generated code across 20+ languages, was released. |
| May 2024 | Doubao foundation models officially launched at the Volcano Engine FORCE conference (Doubao-Pro / Lite, 32K/128K), driving large-model adoption with aggressive pricing — the biggest milestone of the year. |
Experience
| Apr 2025 | XPENG Robotics |
|---|---|
| Jan 2023 | Seed, ByteDance |
| Jun 2021 | ByteDance |
| Sep 2019 | Tencent, focusing on large-scale reinforcement learning for MOBA and FPS video games. |
| Jul 2018 | Momenta, established the reinforcement learning team for L3/L4 autonomous driving. |
Education
| 2017–2018 | Postdoctoral Research Fellow, Harvard University |
|---|---|
| 2014–2017 | Ph.D., The University of Hong Kong |
| 2015–2016 | Visiting Student, National University of Singapore |
| 2009–2013 | B.E., Soochow University |
Selected Publications
-
2026Capek 0.5: An Execution-Centric Vision-Language Model for Embodied IntelligencearXiv preprint arXiv:2608.06756
-
2026R2S-EGO: Dual-Proxy Refinement for Sparse-Capture Real-to-SimarXiv preprint arXiv:2608.06827
-
2026Athena-Brain Technical Report: An Efficient Robot Brain for General Intelligence and Embodied InteractionarXiv preprint arXiv:2607.18985
-
2026Athena-WBC: Capability-Aligned Policy Experts for Long-Tail Humanoid Whole-Body ControlarXiv preprint arXiv:2607.04837
-
2026DeepInsight: A Unified Evaluation Infrastructure Across the Physical AI StackarXiv preprint arXiv:2606.17574
-
2026Engagement Process: Rethinking the Temporal Interface of Action and ObservationarXiv preprint arXiv:2605.11484
-
2026Vigil: Done, But Not Sure — Disentangling World Completion from Self-Termination in Embodied AgentsarXiv preprint arXiv:2605.08747
-
2025Teaching Language Models to Critique via Reinforcement LearningInternational Conference on Machine Learning (ICML)
-
2024Unlock the Correlation between Supervised Fine-Tuning and Reinforcement Learning in Training Code Large Language ModelsarXiv preprint
-
2024Seed-CTS: Unleashing the Power of Tree Search for Superior Performance in Competitive Coding TasksarXiv preprint
-
2024Exploring Mathematical Extrapolation of Large Language Models with Synthetic DataFindings of the Association for Computational Linguistics: ACL
-
2023Mastering Strategy Card Game (Hearthstone) with Improved TechniquesIEEE Conference on Games (CoG)
-
2023Mastering Strategy Card Game (Legends of Code and Magic) via End-to-End Policy and Optimistic Smooth Fictitious PlayarXiv preprint arXiv:2303.04096
-
2017Data-driven Methods towards Learning the Highly Nonlinear Inverse Kinematics of Tendon-driven Surgical ManipulatorsThe International Journal of Medical Robotics and Computer Assisted Surgery
-
2016Motion Planning based on Learning from Demonstration for Multiple-Segment Flexible Soft Robots Actuated by Electroactive PolymersIEEE Robotics and Automation Letters (RA-L)
-
2016Automate Surgical Tasks for a Flexible Serpentine Manipulator via Learning Actuation Space Trajectory from DemonstrationIEEE International Conference on Robotics and Automation (ICRA)