重點導讀
RoboQuest 是用來測試機器人能不能先搜尋、檢查,再做實際測試的評量。
- 表現最好的模型只完成 23.2% 的任務。
- 但單獨測試任務中用到的動作技能時,同一批模型有 72–81% 的成功率。這表示難題不在於機器人會不會動,而在於能不能找齊證據再做決定。機器人常已看到值得檢查的線索,卻太早停止探索或急著行動。
- 要讓機器人更可靠,就得加強追蹤證據、規劃和出錯後的補救能力。
AI 伴隨問答
針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:
elvis@omarsar0
·2 小時前
New benchmark to test general physical agents.
RoboQuest asks robots to search, inspect, and run physical tests before acting. The strongest model solved only 23.2% of tasks, even though the same models completed the underlying motor skills 72–81% of the time in isolation.
The bottleneck was evidence gathering and decision-making.
Agents often saw what they needed to inspect, but stopped exploring or committed too early.
Reliable physical agents need better evidence tracking, planning, and recovery. Great work from @soujanyaporia and colleagues.
Paper: https://t.co/5RjJ3BroWe
推文來源 · X (Twitter)在 X 參與討論