InSight: Self-Guided Skill Acquisition via Steerable VLAs

AI 导读

InSight通过结合视觉-语言模型与可操控VLAs，实现机器人无需人类示范即可自主学习新技能，显著提升任务成功率。

InSight bridges VLMs and VLAs by enabling self-guided skill acquisition without human demonstrations, achieving high success rates in complex tasks.

重点速览

• InSight利用VLM识别VLA缺失的原始技能并进行物理执行 InSight identifies missing primitives for VLAs via VLM and grounds them through robot execution
• 通过成功任务片段强化训练数据，无需完整任务示范 Strengthens training data with successful task segments without full-task demonstrations
• 在扭转、倒水等任务中成功率超90%，远超零样本基线 Achieves 92% twisting and 96% pouring success vs 32%/16% zero-shot baseline
• 支持组合多技能执行，实现复杂任务链操作 Enables multi-skill composition for complex task sequences

一句话：无需人工示范的自主技能学习框架，显著提升机器人任务执行效率。 / Self-guided skill acquisition framework achieves efficient task execution without human demonstrations.

推荐理由 为机器人自主学习提供新范式，降低人工干预成本，具实际应用价值。