How Human Data Becomes Real Robot Capabilities: Ropedia Releases Six Research Findings
人类数据如何变成机器人真实能力?Ropedia重磅发布六项研究成果
During Ropedia Super Week, the Ropedia team released six research findings exploring how human experience can be transformed into training data for robot learning. The research covers cross-robot world prediction, object trajectory recording, first-person reconstruction, programming agent enhancement, bimanual collaboration strategies, and code-driven robot policies. These results demonstrate the diverse applications of large-scale human experience in robot learning and provide new directions for reducing the need for robot-specific data.
The paper presents a series of research contributions focused on leveraging human experience for robot learning, including world modeling, data annotation, and policy learning. It highlights the use of first-person data to improve robot perception and task execution, with results showing improved performance across multiple benchmarks and simulated environments.
Source: Chinese robotics — Dataset and collection discovery · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
human first-person video hours
≈10,000 hours
View original evidence
为了训练这一模型,研究团队构建了 Xp-Data,整合约 1 万小时人类第一视角视频和 3000 小时机器人视频 ,统一相机轨迹、运动骨架和任务描述等多模态标注。
Open source S11robot video hours
3,000 hours
View original evidence
为了训练这一模型,研究团队构建了 Xp-Data,整合约 1 万小时人类第一视角视频和 3000 小时机器人视频 ,统一相机轨迹、运动骨架和任务描述等多模态标注。
Open source S11PSNR score
21.98
View original evidence
在覆盖八种本体的跨本体评测中,同一组 Ropedia W0 权重无需针对目标评测数据额外训练,就取得了 21.98 dB 的 PSNR,高于使用相同机器人混合数据适配的 Cosmos3-Nano(19.02 dB),展现出跨本体的世界预测能力。
Open source S16Reported success rate
56.82%
View original evidence
在 LIBERO-Plus 基准上,Ropedia W0 初始化版本的成功率由 50.06% 提升至 56.82% ,并在相机、光照、背景等七类环境扰动下均取得改善。
Open source S23Reported success rate
50.06%
View original evidence
在 LIBERO-Plus 基准上,Ropedia W0 初始化版本的成功率由 50.06% 提升至 56.82% ,并在相机、光照、背景等七类环境扰动下均取得改善。
Open source S23
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
为了训练这一模型,研究团队构建了 Xp-Data,整合约 1 万小时人类第一视角视频和 3000 小时机器人视频 ,统一相机轨迹、运动骨架和任务描述等多模态标注。
Open source S11
在覆盖八种本体的跨本体评测中,同一组 Ropedia W0 权重无需针对目标评测数据额外训练,就取得了 21.98 dB 的 PSNR,高于使用相同机器人混合数据适配的 Cosmos3-Nano(19.02 dB),展现出跨本体的世界预测能力。
Open source S16
在 GR-1 评测中,使用 W0 预训练权重的模型, 仅需 1% 的机器人后训练数据,预测质量就超过了使用全量数据的通用视频生成模型初始化版本 。
Open source S18
在规划任务中,机器人可以先通过模型预测不同候选动作的结果,再选择更合适的动作执行,真机任务完成进度由随机选择时的 26% 提升至 55% 。
Open source S21
在 LIBERO-Plus 基准上,Ropedia W0 初始化版本的成功率由 50.06% 提升至 56.82% ,并在相机、光照、背景等七类环境扰动下均取得改善。
Open source S23
Source:Chinese robotics — Dataset and collection discovery · finance.sina.com.cn