OM-1: One Model, One Data Interface, Any Body (Reward AI)
博客要点 (中文)
Reward AI 在 2026-09-14 发布的 OM-1 博客介绍跨机身机器人基础策略。口号：One Model, One Data Interface, Any Body。能力宣称：桌面臂 / 工业臂 / 人形机身 zero-shot 迁移；新任务 <30 分钟人类演示数据上手；覆盖调酒、叠衣服、包装、拔网线等长程接触丰富任务的实时控制 + 高频 RL 控制层。关键数据边界：训练数据来自穿戴式 Omnibody Hand（7-DoF，含触觉 / 接近 / 手内相机 / 力 / 电磁跟踪，沿 Stanford DexCap 工作延伸），不含 teleoperation、无 on-robot 预训练数据写入 OM-1——与 VLA「机器人轨迹预训练」路线最大的边界区别。本机未复跑；写具身稿时把「数据里是否有 robot experience」做成硬边界。这与手机 agent「点屏 MCP vs AppFunctions」属不同 embodiment 层——本文赌的是「人类动作接口是否足以替代 per-robot 数据飞轮」。
Key claims (English)
Reward AI's OM-1 blog (2026-09-14) introduces a cross-embodiment robotics foundation policy under the slogan 'One Model, One Data Interface, Any Body'. Capability claims: zero-shot transfer across desktop arms, industrial arms, and humanoids; new tasks picked up with under 30 minutes of human demonstration; long-horizon contact-rich tasks (mixing drinks, folding laundry, packing, unplugging cables) under real-time control plus a high-frequency RL control layer. Critical data boundary: training data comes from the wearable Omnibody Hand (7-DoF with tactile / proximity / in-hand camera / force / EM tracking, building on Stanford DexCap), with no teleoperation and no on-robot pretraining data written into OM-1 — the sharpest contrast to VLA-style 'robot trajectory pretraining'. The author has not re-run it locally; the hard boundary for embodied-AI writeups is 'does the data contain robot experience'. This sits in a different embodiment layer than the phone-agent 'tap-screen MCP vs AppFunctions' debate; the bet is that a human-motion interface can substitute for a per-robot data flywheel.
Obsidian 证据摘录
「口号：One Model, One Data Interface, Any Body。数据：穿戴式 Omnibody Hand（7-DoF，触觉 / 接近 / 手内相机 / 力 / 电磁跟踪；DexCap 系 Stanford 工作延伸）；无 teleoperation、无 on-robot 预训练数据写入 OM-1。能力宣称：桌面臂、工业臂、人形 zero-shot 迁移；新任务可用 <30 分钟人类演示数据；长程接触丰富任务（调酒、叠衣服、包装、拔网线等）实时控制 + 高频 RL 控制层。」——Hermes 定时任务/X-每日简报/2026-09-14-X-Hot-Brief_AtriaDawn-OM1-GlassImaging.md L53-55