Xu Mengdi / 徐梦迪

Overview

Xu Mengdi is a robot-learning researcher and assistant professor at 清华大学交叉信息研究院. The bounded source traces her path from vehicle engineering and biomimetic and snake-robot hardware into learning algorithms at Carnegie Mellon and postdoctoral work with Fei-Fei Li and Jiajun Wu at Stanford.

Current Profile

Xu’s research identity centers on robots that continue learning in changing real environments instead of being fully defined by pretraining data. She connects meta-learning, decision transformers, in-context learning, world models, multi-source robot data, and deployment feedback to a two-part view of generalization: a robot needs strong prior capability before deployment and fast adaptation once novelty appears. Her methodological stance also emphasizes choosing a consequential long-term problem, defining success at the task level, and treating hardware, models, data, evaluation, safety, and deployment as one system.

Key Characteristics

• Studies rapid robot adaptation through context, demonstrations, corrections, failures, and test-time experience.
• Distinguishes genuine new-task learning from interpolation within a familiar training distribution.
• Favors world-model research for learning task-independent change while allowing complementary VLA execution layers.
• Treats data collection as capability-driven and stage-specific rather than a contest for the largest undifferentiated hour count.
• Uses unseen-task success, safety boundaries, and real-time behavior as stronger evaluation targets than held-out loss alone.
• Frames research identity as sustained attention to a small number of important questions.

Evidence

• Adaptation research: “我看到了 Scaling Law 的信号”｜对谈清华叉院助理教授徐梦迪：具身智能、世界模型、真正的泛化 connects Xu’s Prompt Decision Transformer, Hyper Decision Transformer, and Algorithm Distillation work to rapid adaptation and learning from execution history.
• Generalization and models: “我看到了 Scaling Law 的信号”｜对谈清华叉院助理教授徐梦迪：具身智能、世界模型、真正的泛化 records her two-part generalization frame and preference for task-independent world modeling over enumerating every task.
• Data and deployment: “我看到了 Scaling Law 的信号”｜对谈清华叉院助理教授徐梦迪：具身智能、世界模型、真正的泛化 assigns real-robot, simulation, UMI, and human data different roles and describes current dual-arm correction and demonstration experiments.
• Evaluation and research method: “我看到了 Scaling Law 的信号”｜对谈清华叉院助理教授徐梦迪：具身智能、世界模型、真正的泛化 distinguishes test loss from success, stresses embodied safety and latency, and links independent research to problem definition and long-term focus.

Qualifications

The profile is bounded to one interview and is not a complete publication record, biography, or independent assessment of Xu’s lab. Technical results, observed scaling signals, project performance, and three-to-five-year forecasts are reported without full experimental protocols or audited comparisons.

What Changed

• Created the profile around Xu’s current synthesis of robot generalization, in-context learning, data, and evaluation.

Relationships

• Tsinghua Institute for Interdisciplinary Information Sciences / 清华大学交叉信息研究院 - current academic institution and lab-building setting.
• Fei-Fei Li - postdoctoral mentor associated with vision-driven problem definition and task-success evaluation.
• Robot In-Context Learning / 机器人上下文学习 - central capability target across Xu’s research trajectory.
• World Models - preferred route for learning task-independent world change.
• Capability-Driven Robot Data Design / 能力反推机器人数据 - data-selection principle articulated in the interview.
• Problem Definition In Research - methodological shift Xu associates with becoming an independent researcher.