RAX: Scaling Online Exploration for Robot Agents
机器人的o1时刻来了！哈佛、伯克利、英伟达推出探索Scaling Law，五大基准平均领先37.4个点
Researchers from Harvard, NVIDIA, UC Berkeley, MIT, and Stanford introduce RAX, a robot learning framework that enables agents to autonomously explore and accumulate skills through online exploration. RAX achieves a 37.4% average improvement over previous state-of-the-art models across five benchmarks, and demonstrates zero-shot performance on real robots without human demonstrations. The system uses code-as-policy to convert exploration into reusable skills, enabling cross-embodiment learning and deployment.
The paper introduces RAX, a framework for robot agents that enables continuous self-improvement through online exploration. It demonstrates significant performance gains across multiple benchmarks without requiring human demonstrations, marking a step toward scalable, autonomous robot learning.
Source: Chinese robotics — Company news 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
Reported duration
<2 hours
View original evidence
真实机器人上，它在没有任何人工动作示范的情况下，仅仅通过不到 2 小时的自主探索，就能熟练掌握并快速执行精密插孔、柔性物体操作、关节物体操作，以及长程移动机器人全身控制等任务。
Open source S14points
37.4
View original evidence
在 RoboDojo、BEHAVIOR-1K、HumanoidArena、DexJoCo 和 LIBERO-Pro 五大仿真基准上，RAX 相对以往最强的 SOTA 模型平均领先 37.4 个百分点，刷新多项 SOTA。
Open source S13
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
真实机器人上，它在没有任何人工动作示范的情况下，仅仅通过不到 2 小时的自主探索，就能熟练掌握并快速执行精密插孔、柔性物体操作、关节物体操作，以及长程移动机器人全身控制等任务。
Open source S14
在 RoboDojo、BEHAVIOR-1K、HumanoidArena、DexJoCo 和 LIBERO-Pro 五大仿真基准上，RAX 相对以往最强的 SOTA 模型平均领先 37.4 个百分点，刷新多项 SOTA。
Open source S13
Source：Chinese robotics — Company news discovery · finance.sina.com.cn