Skip to content
Source
Machine Heart — Robotics on WeChat·· 2 days agoSignalEditorial score85

CoRL 2026｜Making Robots More Agile: Starting with Learning to Walk

CoRL 2026｜让机器人更会动，从学会「起步」开始

Summary

This paper presents LeaP, a learnable source prior for generative robot policies, which replaces fixed standard Gaussian distributions with state-dependent random initialization. LeaP jointly learns the mean and variance of an initial Gaussian distribution, improving action generation in robotic tasks. The method achieves an average success rate of 81.6% in simulation and 80.0% on real robotic arms, outperforming existing baselines like A2A and VITA. The paper is accepted at CoRL 2026, and the code is open-sourced.

Full article

You are reading the complete RoboSignal summary. The publisher’s full article is available at the original source.

Read full article at source

mp.weixin.qq.com · Opens in a new tab; source language may differ.

Editorial context

This paper introduces LeaP, a learnable source prior for generative robot policies, which improves action generation by incorporating state-dependent random initialization. The method outperforms baselines in both simulation and real-world robotic tasks, demonstrating the effectiveness of modeling uncertainty in action generation.

What the source reports

Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.

Reported numbers

  • Reported success rate

    81.6%

    View original evidence
    在 RoboTwin 的 15 项仿真操作任务上，其平均成功率达到 81.6 % ，较采用相同编码器与生成器的标准高斯基线提升 25.5 个百分点。
    Open source S8

  • Reported success rate

    80%

    View original evidence
    LeaP 平均成功率为 80.0% ，高于 A2A 的 68.3%、VITA 的 56.7% 和 NoPrior 的 46.7%；相较 NoPrior 提升 33.3 个百分点。
    Open source S17

  • Reported success rate

    85.3%

    View original evidence
    仅用本体感知的 LeaP 平均成功率达到 85.3%，视觉及三种视觉与状态融合方案则为 59.3%—70.3%。
    Open source S19

  • Reported success rate

    91%

    View original evidence
    在打开笔记本电脑任务的实验中，LeaP 训练 1000 轮达到 91% 成功率，超过四个主要基线训练 3000 轮后的表现，展示了改善收敛效率的潜力。
    Open source S15

  • Reported success rate

    76.7%

    View original evidence
    在扩散桥中，相比采用先验均值的确定性起点，完整 LeaP 分布将成功率从 68.7% 提升至 76.7%。
    Open source S24

  • Reported success rate

    59.7%

    View original evidence
    仅通过流匹配损失训练先验，平均成功率为 59.7%，高于无先验基线的 47.7%，但仍低于完整模型的 85.3%。
    Open source S22

  • Reported success rate

    78%

    View original evidence
    更关键的是一组三方对照：仅使用预测均值时，成功率为 78.0%；加入固定标准差为 1 的高斯噪声后，反而降至 62.7%；联合学习均值与状态自适应方差则达到 85.3%。
    Open source S21

  • success_rate

    68.3%

    View original evidence
    LeaP 平均成功率为 80.0% ，高于 A2A 的 68.3%、VITA 的 56.7% 和 NoPrior 的 46.7%；相较 NoPrior 提升 33.3 个百分点。
    Open source S17

What remains unknown

Not established in the collected evidence: 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.

在 RoboTwin 的 15 项仿真操作任务上，其平均成功率达到 81.6 % ，较采用相同编码器与生成器的标准高斯基线提升 25.5 个百分点。

Open source S8

实验结果与分析 团队在 RoboTwin 的 15 项双臂操作任务上检验这一设计，涵盖抓取放置、工具使用与双臂协同。

Open source S13

在 Franka Research 3 的抓取方块、关闭盒子、抓取并放置沙袋三项任务中，每项使用 100 条遥操作演示，并随机设置物体位置测试 20 次。

Open source S16

实验在抓取不同瓶子、打开笔记本电脑、交接方块三项 RoboTwin 任务上进行，每项任务评估 100 次。

Open source S18

LeaP 平均成功率为 80.0% ，高于 A2A 的 68.3%、VITA 的 56.7% 和 NoPrior 的 46.7%；相较 NoPrior 提升 33.3 个百分点。

Open source S17

仅用本体感知的 LeaP 平均成功率达到 85.3%，视觉及三种视觉与状态融合方案则为 59.3%—70.3%。

Open source S19

在打开笔记本电脑任务的实验中，LeaP 训练 1000 轮达到 91% 成功率，超过四个主要基线训练 3000 轮后的表现，展示了改善收敛效率的潜力。

Open source S15

同一先验使流匹配的平均成功率从 47.7% 提升至 85.3%；在扩散桥中，相比采用先验均值的确定性起点，完整 LeaP 分布将成功率从 68.7% 提升至 76.7%。

Open source S24

仅通过流匹配损失训练先验，平均成功率为 59.7%，高于无先验基线的 47.7%，但仍低于完整模型的 85.3%。

Open source S22

更关键的是一组三方对照：仅使用预测均值时，成功率为 78.0%；加入固定标准差为 1 的高斯噪声后，反而降至 62.7%；联合学习均值与状态自适应方差则达到 85.3%。

Open source S21

论文标题：Where Should Action Generation Begin? A Learnable Source Prior for Generative Robot Policies 论文链接：https://arxiv.org/abs/2606.17408 项目仓库：https://github.com/SEU-VIPGroup/LeaP 项目主页：https://daimeipo.github.io/LeaP/ 图 1｜三种动作生成起点：标准高斯噪声、由观测确定的起点，以及 LeaP 学习的条件概率分布。

Open source S10

Implications for data suppliers

RoboSignal interpretation and collection questions, not statements of buyer demand.

  • Confirm the required data type and collection setting with the buyer; this source does not establish a complete collection specification.
  • Compare the reported units and scope before using these quantities in a budget. Recording hours, sensor-hours and trajectories are different measures.
  • Validate demand and acceptance criteria with a buyer before scaling. Publication, popularity and a research result do not establish a purchase commitment.

Source：Machine Heart — Robotics on WeChat · mp.weixin.qq.com

#Research#Robot policies#Imitation learning#Reinforcement learning#Embodied perception#Open source