Figure releases Helix 2.5 model for home tasks
人形机器人可“自主做家务”，Figure 发布 Helix 2.5 模型
For users interested in robotics and AI, this update shows significant progress in making humanoid robots more adaptable and capable of performing real-world tasks without extensive retraining.
Figure, a humanoid robot company, released the Helix 2.
5 model, which allows robots to autonomously perform household chores without extra training.The company tested the model in 30 homes in the San Francisco Bay Area, where the robots started working immediately upon arrival.
The model was pre-trained on the Index dataset of human behavior, and its single base model was adapted to three behaviors: tidying living rooms, folding towels, and making beds.
Compared to its predecessor, Helix 02, the new model achieved a 56% task success rate with half the specialized data, demonstrating improved zero-shot generalization to new environments.
人形机器人可“自主做家务”，Figure 发布 Helix 2.5 模型
IT之家 9 月 18 日消息，人形机器人企业 Figure 今日发布了 Helix 2.5 模型， 能让机器人“自主做家务” 。 IT之家从 Figure 官方介绍获悉，Figure 在美国旧金山湾区租了 30 套房子，机器人没有进行额外训练，到了地方就开始自己收拾房子。 Helix 2.5 的构建旨在回答一个问题 —— 类人生物能否进入一个它从未见过的家，并立即自主工作？ 为此，Figure 在全球人类行为数据集 Index 上预训练了 Helix 2.5。基于这个单一的基础模型，机器人产生了三种行为： 整理客厅、折叠毛巾和整理床铺 。然后 Figure 带着机器人进入了湾区 30 户家庭，之前没有收集任何数据。 在测试中，一个单一的 Index 预训练基础模型被调整为三种不同行为，涵盖运动、刚性和可变形操作、双手协调以及主动感知。 根据实测数据，Index 预训练显著提高了人形机器人零样本全身泛化能力，在任务专用数据、架构、训练和评测均保持不变的情况下， 该变量使得任务成功率从 8% 提升到了 56% ；与前代神经网络 Helix 02 相比，Helix 2.5 只用了一半的任务专用数据，就将该行为泛化到了 30 个前所未见的新环境。 Figure 表示，Helix 2.5 首次证明，全身智能可以从人类经验中学习并转移到新的场景中，而不用每次都进行重建。
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, humanoid robots, AI models.
Source attribution: curated from IT之家.
