Figure releases Helix 2.5 robot model

To test its real-world capabilities, the company rented 30 homes in the San Francisco Bay Area and sent its robots into unvisited houses without prior data collection.

The robots were tasked with cleaning living rooms, folding towels, and making beds.The model demonstrated a key qualitative ability: self-correction after errors, such as repositioning or adjusting its approach when encountering obstacles or mistakes.

This self-recovery during long, multi-step tasks is attributed to the Index pre-training.

Figure 发布 Helix 2.5 机器人模型 普通人走进一个陌生的家里，马上就能动手干活，那么机器人呢？ 为了测试该模型的真实能力，Figure 在旧金山湾区租下 30 栋团队和机器人从没进去过的房子 事先不在里面采集任何数据，让机器人直接进去收拾客厅、叠毛巾、铺床。 测试机器人的泛化能力... 在长时间、多步骤的任务里，Helix 2.5 表现出的一项定性能力是全身自我纠错：遇到失误或动作受限时，机器人能够退后一步重新调整站姿、换个角度，或者绕着整张床走过去纠正一个被角。 Figure 认为，这种在出现失误后自行恢复并继续推进任务的能力，是 Index 预训练带来的一个重要效果...

This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Figure, Helix 2.5, robots.

Source attribution: curated from 小互 (@imxiaohu).