Sim2Real

Sim2Real is the robotics route of training or testing behavior in simulation and transferring it to real robots. 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 argues that robot manipulation cannot skip Sim2Real if the goal is reliable operation on unseen objects and unseen environments.

The source’s version of Sim2Real is not simply prettier virtual worlds. It emphasizes Structured 3D Robot Data, physical consistency, GPU-parallel environments, hardware-specific noise, and bottom-level manipulation skills that can survive transfer to the real robot body.

Key Claims

• Simulation must be physically useful, not merely visually realistic.
• The simulator and hardware have to co-evolve because motor noise, response curves, gripper structure, and boot-time differences can affect transfer.
• Real robot data remains necessary, but it is too scarce and expensive to be the only source of open-world manipulation learning.
• Robotics Simulation Evaluation and Embodied Data Pyramid become practical infrastructure when Sim2Real is treated as training, evaluation, and feedback rather than as a single transfer trick.

Connections

• 速度科技 / Sudu Technology and Han Zheng / 韩正 - source company and guest.
• ManiSkill, ShapeNet, and PartNet / PartNet Mobility - project and data-set context.
• Open-World Robot Manipulation and Layered Robot Architecture - capability target and architecture route.
• Real Robot Data Strategy - route the source qualifies rather than rejects.