Jim Fan 团队开源机器人数据引擎 Flora，称成本比 Astra 低 125 倍

摘要

Jim Fan 团队开源了 Flora，一套用于解析物理世界、把原始视频转成训练信号的开源模型，并公开模型权重。Flora 已在其内部数据引擎中运行，每天在 500 块 GPU 上为 20 万小时视频打标。它提供带时间戳的高质量描述用于筛选人类第一视角与机器人数据、用于质量控制和强化学习后训练的密集进度奖励，以及面向真实世界基准的评分规则。

摘要由模型根据原文撰写，以原文为准。

We built a robotics data engine with Astra-level performance at 125x lower cost. And we are open-sourcing the weights today! Introducing Flora, a suite of open models to parse the physical world and turn raw footage into rich training signals.

Flora already powers our in-house data engine, labeling 200K hours of videos a day on 500 GPUs. We created it to scale the frontier for robot learning. Now you own the weights too.

Flora gives you:

High-quality, timestamped captions to curate and filter human egocentric and robot data;
Dense progress rewards for quality control and RL post-training;
Rubrics scoring for real-world benchmarking

Try it today, links in thread:

DrJimFan. We built a robotics data engine with Astra-level performance at 125x lower cost. And we are open-sourcing the weights today! Introducing ...[EB/OL]. (2026-10-21)[访问日期]. https://x.com/DrJimFan/status/2112597587987689912.

We built a robotics data engine with Astra-level performance at 125x lower cost. And we are open-sourcing the weights today! Introducing ...

摘要

正文 · 原文

引用原文