From 12 to 200 Demonstrations: How ACT, Diffusion Policy, SmolVLA and $π_{0.5}$ Scale on a Low-Cost Arm

📄 arXiv: 2610.37170v1 📥 PDF

作者: Simon Cowling

分类: cs.RO

发布日期: 2026-10-27

备注: 31 pages, 5 figures, 19 tables

💡 一句话要点

通过多样化示例提升低成本机械臂的模仿学习效果

🎯 匹配领域: 支柱二：RL算法与架构 (RL & Architecture)

关键词: 模仿学习 机械臂 数据稀缺 策略优化 机器人技术

📋 核心要点

  1. 现有模仿学习策略在数据稀缺时表现不佳，尤其是在复杂任务中，如抓取和放置。
  2. 本研究通过训练多种模仿学习策略，探索不同数量的示例对学习效果的影响，旨在优化低成本机械臂的操作能力。
  3. 实验结果表明，ACT在数据量增加时成功率显著提升，而$π_{0.5}$在数据稀缺时表现优异，提供了对比分析和策略优化的依据。

📝 摘要（中文）

本研究训练了四种广泛使用的模仿学习策略：ACT、Diffusion Policy、SmolVLA和$π_{0.5}$，在不同数量的远程操作示例（12、25、50、100和200个）上进行评估。实验结果显示，ACT在数据量增加时表现显著提升，而$π_{0.5}$在数据稀缺时表现最佳。SmolVLA和Diffusion Policy在启动失败方面存在问题，影响了整体表现。研究还分析了不同策略在不同阶段的失败情况，并探讨了数据量与训练效果之间的关系。

🔬 方法详解

问题定义：本研究旨在解决低成本机械臂在模仿学习中对数据量的依赖性，尤其是在复杂任务（如抓取和放置）中的表现不足。现有方法在数据稀缺时往往无法有效学习，导致成功率低下。

核心思路：通过对四种模仿学习策略进行系统训练，分析不同数量的示例对学习效果的影响，探索如何在数据稀缺的情况下优化策略表现。

技术框架：研究采用固定的训练方案对四种策略进行训练，使用不同数量的示例进行评估。每种策略在物理机械臂上进行测试，记录每个阶段的成功与失败情况。

关键创新：本研究的创新点在于系统性地比较了多种模仿学习策略在不同数据量下的表现，揭示了数据量与策略成功率之间的关系，尤其是ACT和$π_{0.5}$的表现差异。

关键设计：训练过程中，保持训练长度固定，确保数据量与训练次数的变化一致。采用不同的微调策略（如低秩适应与动作专家微调）来优化$π_{0.5}$的表现，并分析其对成功率的影响。实验中还考虑了种子间的变异性与噪声的影响。

📊 实验亮点

实验结果显示，ACT在12个示例时成功率为2.5%，而在100个示例时提升至92.5%。$π_{0.5}$在数据稀缺时表现最佳，12个示例成功率为40%，而在25个示例时为70%。SmolVLA和Diffusion Policy的成功率均未超过50%，且存在启动失败的问题。

🎯 应用场景

该研究的成果可广泛应用于机器人操作、自动化制造和智能家居等领域。通过优化模仿学习策略，可以提升低成本机械臂在复杂任务中的执行能力，降低对高质量数据的依赖，推动智能机器人技术的普及与应用。

📄 摘要（原文）

We trained four widely used imitation-learning policies, ACT, Diffusion Policy, SmolVLA and $π_{0.5}$, on nested subsets of 12, 25, 50, 100 and 200 teleoperated demonstrations of one pick-and-place task with distractor objects, using one fixed training recipe per policy. Every trained model was then run on a physical SO-101 arm on a fixed set of object placements: 1,340 scored episodes in 126 blocks. ACT, trained from scratch, rose from 1 success in 40 episodes at 12 demonstrations to 37 in 40 at 100. $π_{0.5}$ with its action expert fine-tuned was the best policy when data was scarce (16 of 40 at 12 demonstrations, 28 of 40 at 25) and then stayed near 70%; from 100 demonstrations ACT was ahead, by a margin this study does not fully resolve. SmolVLA and Diffusion Policy peaked near 50%, and both were affected by a failure to start: SmolVLA's arm never started moving in 47% of episodes, and Diffusion Policy hardly ever started until an idle lead-in was trimmed from the demonstrations. For $π_{0.5}$ the fine-tuning recipe mattered more than the amount of data: low-rank adaptation succeeded in 4 of 110 episodes, against 130 of 200 for action-expert fine-tuning. From the joint and camera recordings we place every episode on a four-stage ladder (started, reached the ball, carried it, delivered it) and show that for ACT more data moves failures from reaching, to grasping, to none, while the failures $π_{0.5}$ retains are in grasping. We also measure how weak a ten-episode block is as a measurement (the same checkpoint tested twice differs by 2 successes in 10 on average), separate seed-to-seed variation from that noise, and audit the human scoring against the recordings. Training length was fixed within each recipe, so the amount of data and the number of times it was seen change together; we state which conclusions survive that confound and which do not.