AntiGrounding: Executable Robot Trajectories as Visual Prompts for VLM-Guided Manipulation

AI Digest

文章提出AntiGrounding框架，通过可执行轨迹作为视觉提示，结合VLM评估实现机器人操作，提升任务成功率与物理合理性。

AntiGrounding uses executable robot trajectories as visual prompts for VLM-guided manipulation, enhancing task success through multi-view assessment and digital twin validation.

Key points

• 双几何-视觉轨迹界面实现动作选择与指令评估 Dual geometric-visual trajectory interface for action selection
• 多视角VQA评分体系评估轨迹安全性与效率 Multi-view VQA scoring system for trajectory evaluation
• 数字孪生验证机制确保真实机器人执行安全 Digital twin validation ensures real-world safety
• 单GPT-6 Astra评估器实现8项任务80%成功率 Single evaluator achieves 80% success across 8 tasks

Takeaway: 轨迹可视化评估与数字孪生验证显著提升机器人操作可靠性。 / Trajectory visualization and digital twin validation enhance robotic manipulation reliability.

Why it matters 提供机器人操作新范式，对工业自动化落地有实践指导价值。