Загрузка видео...
Не удалось загрузить видео
From Failures to Learnable Worlds Policy-Conditioned Environment Generation for Navigation Agents paper:
17,068 просмотров • 1 месяц назад •via X (Twitter)
Комментарии: 6

lewington1 месяц назад
i think it’s just a personalised curriculum for embodied agents

陈放Frank🕊️1 месяц назад
「导航 agent 失败 → 反过来生环境再训练」这招特别狠——把失败率当数据用而不是砸更多算力,之前 Voyager 那条线就在做这件事,但生成环境的可控性一直卡。 policy-conditioned 比 unconditional 难不少,主要卡在 conditional signal 怎么拿到(动作 label + 失败 case 的对齐)。如果这个方向真跑通,对 Embodied Turing Machines 那类 self-improve 也是直接补一块——agent 可以自己做 curriculum。

LongTian1 месяц назад
failing forward is the real curriculum here, and conditioning on the policy makes the generated worlds actually bite back

Daniel Priscu1 месяц назад
building the next environment from where the policy fails beats burning compute on levels it already solved

Broadstreet1 месяц назад
⏫

Elio科技迷(互)1 месяц назад
turning failed runs into new training worlds is a neat trick
