扁平化科技风机械臂手部抓取流程插画素材

Draw a single academic-style schematic diagram (ICRA/RSS/CoRL) showing a three-stage pipeline for dexterous grasping with Mocap, ECN, and RL.Left: "MoCap to Physics-Consistent Dataset".Raw MoCap hand trajectory (non-physical) → Retarget + RL reconstruction → physics-consistent hand-object interaction.Annotate output as "Physical Contact Ground Truth (s_t^i, df_t^i)".Middle: "ECN Training".Input: raw MoCap hand + object point cloud (non-physical).Supervision: physical contact ground truth from left (dashed arrow).ECN outputs predicted contact state and force change rate.Highlight mismatch between non-physical input and physical supervision.Right: "ECN-Guided RL Grasping".RL policy takes state input.Frozen ECN provides auxiliary contact prediction as observation or reward guidance.RL outputs optimized dexterous grasp motion. Style: clean, minimalistic academic figure, clear boxes and arrows, readable labels, color-coded modules, emphasize data flow and supervision.