Podcast Atlas
Source note Episode guide Original audio Topics: Technology

“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

Summary

This 十字路口Crossing episode interviews 徐梦迪, an assistant professor at 清华大学交叉信息研究院, about robot learning beyond a fixed pretraining distribution. Its central claim is that genuine generalization combines strong priors with fast in-context adaptation from demonstrations, corrections, bodily motion, and failures; World Models and VLA systems can be complementary, while data sources should be assigned by the capability and training stage they best support. The episode also strengthens the wiki’s caution around embodied scaling and evaluation: lower held-out loss is an early signal, not a scaling law for useful robotics unless unseen-task success rises reliably under safety and real-time constraints.

Key Claims

  • Genuine robot generalization requires both broad pretrained knowledge and the ability to learn quickly after deployment; interpolation within familiar object positions is not equivalent to learning a new task.
  • Robot In-Context Learning / 机器人上下文学习 aims to adapt behavior from demonstrations, language corrections, body motion, and suboptimal or failed trajectories without relying entirely on parameter updates.
  • Prompt Decision Transformer, Hyper Decision Transformer, and Algorithm Distillation represent successive attempts to make model parameters encode a learning procedure rather than only memorized task behavior.
  • World Models target task-independent regularities of how the world changes, while VLA models map perception and instructions toward action; the episode treats understanding/prediction and execution as potentially complementary.
  • Capability-Driven Robot Data Design / 能力反推机器人数据 assigns real-robot teleoperation, simulation, UMI-style collection, and human data different roles based on the desired capability and training stage rather than declaring one universal best source.
  • Robot Deployment Data Loop / 机器人部署数据闭环 depends on a sufficiently capable base model: users will keep teaching and correcting a robot only when limited interaction produces visible improvement.
  • Held-out loss reportedly improves as pretraining data grows from roughly one hundred thousand to one million hours, but loss improvement is not task-success scaling, especially in long-horizon contact-rich tasks.
  • Embodied evaluation should measure unseen-task success, failure modes, safety boundaries, and latency because a robot’s body turns errors into physical consequences and makes slow inference operationally limiting.
  • Long-term research identity can emerge from sustained work on a consequential problem; independent researchers must define worthwhile questions as well as solve them.

Key Quotes

“真正的泛化” — the episode’s label for combining prior knowledge with rapid learning in new environments.

“损失和成功率并不直接挂钩” — the core qualification on interpreting an embodied-AI scaling signal.

“学习如何学习” — the capability the episode connects to repeated adaptation across homes and service settings.

Connections

Contradictions

  • The episode does not overturn the existing Robot Scaling Claim Caution / 幂律现象与 scaling law 谨慎; it adds a somewhat stronger observed signal while preserving the same evidentiary boundary. Held-out-loss improvement across source-reported data scales does not establish a power law for unseen-task success.
  • The preference for world models over task-enumerating VLA training is a research judgment, not comparative proof that one route dominates. The source itself leaves room for world-model/VLA complementarity.
  • Claims about data-hour scale, current industry results, benchmark adequacy, future inflection timing, and the guest’s team experiments remain source-scoped because the episode supplies no audited curves, protocols, or cross-model comparisons.