Layered Robot Architecture

Layered robot architecture is the episode’s route distinction between high-level reasoning/planning and low-level physical manipulation. In E244｜端到端vs上下分层：机器人路径之争，正在转向？, Han Zheng / 韩正 predicts that this layered structure may return to the mainstream as teams hit the limits of scarce real robot data and narrow end-to-end imitation.

146. 对Physical Intelligence柯丽一鸣4小时访谈：Pi的开源模型研究，机器人的江湖、族谱与主角 adds a bridge from K’s background: his doctoral environment trained him in traditional robotics concerns such as planning, control, and real-machine performance, even as his later work moved toward imitation learning, Robot Reinforcement Learning, and VLA-style robot policies. The source therefore treats layered and learned methods as historical influences to combine, not only rival camps.

The source defines the upper layer as understanding the environment, decomposing the task, and choosing a plan. The lower layer is concrete manipulation such as opening a door, grabbing a bottle, unscrewing a cap, pouring water, or inserting an object. The source’s nuance is that deployment can still be end-to-end at the edge while pretraining uses interpretable intermediate structure.

Key Claims

• Layering is not the same as brittle rule stitching; it can include learned geometry, material, kinematics, object-part, and future-state representations.
• Reliable short manipulation skills are prerequisites for long-horizon tasks because failures compound across steps.
• Physical Intelligence and Generalist are framed as stronger in upper-level reasoning or robot-brain work, while 速度科技 / Sudu Technology emphasizes bottom-level manipulation.
• The architecture works best when tied to Sim2Real, hardware/software co-design, and Structured 3D Robot Data.
• K’s account adds that learned robot policies still inherit real-machine discipline from traditional robotics: the route only matters if the robot can physically do the task.

Connections

• Vision Language Action Models and World Model VLA Fusion - adjacent model architectures.
• ManiSkill - short-skill and task-composition context.
• Open-World Robot Manipulation - evaluation bar for the lower-level skills.
• Embodied AI Value Chain - business question of who owns the brain, body, data, and developer platform.
• K, Robot Reinforcement Learning, and Robot Experience Data - experience-learning route that intersects with traditional robotics constraints.