WorldFoundry 将真实示范转化为可执行、可追溯的仿真实验:以四路录像和配套关节反馈为输入,结合厂家机器人模型、夹爪 CAD 与估计杯壳几何,分别构建物体、场景和任务。原反馈按原时间轴输入 Isaac Sim,由 PhysX 在 240 Hz 下求解接触与杯子运动;独立评估器同时检查动作执行、接触支撑、释放稳定性及相对真实参考的轨迹差异。每次失败都回溯到具体几何、参数或控制候选,以新版本复验。当前原动作在 1 mm 有效杯壁模型中仍出现红杯侧倒,尚未实现 1:1 物理复现;同一场景中重新生成的三条局部控制变体通过全部 18 项任务检查,形成 54,216 条状态—动作—下一状态转换。该案例把视觉对齐、物理任务成功与真实一致性分开度量,使后续参数辨识和训练数据筛选具有明确证据。
English abstract
WorldFoundry turns a recorded demonstration into an executable and traceable simulation experiment. Four camera streams and companion joint feedback are combined with a manufacturer robot model, gripper CAD, and estimated hollow-cup geometry through object, scene, and task harnesses. Recorded feedback is applied on its original clock, while Isaac Sim / PhysX resolves contact and free cup motion at 240 Hz. An independent evaluator checks task execution, support, release stability, and discrepancy from the available real reference. Failures guide explicit revisions of geometry, parameters, or control candidates, each evaluated as a new version. The recorded-motion attempt in the 1 mm effective-wall scene ends with the red cup tipped over; one-to-one physical reproduction is not yet achieved. Three newly generated local control variants in the same scene pass all 18 task gates and yield 54,216 state–action–next-state transitions. The experiment separates visual alignment, physical task validity, and agreement with reality to support subsequent identification and data admission.
实线为本案例已执行的数据链与候选修订循环;底部虚线区域列出仍需实测的参数辨识。一次原生执行可渲染多视角,视角数量不增加行为条数。 PDF · SVG · 可编辑源代码
Object / Scene / Task harness
Harness 是组织输入、执行器、验证器与证据的程序框架。本案例由任务专用后端接入这套分层组织;通用 scene 接口目前只处理规范化刚体摆放,双臂叠杯并非任意场景、任意资产的通用策略。