Skip to content

论文arXiv 2611.07441今天

低成本物理仿真评测揭示动作分块策略在稀疏数据下失效 ​

本文由 AI 根据公开内容整理,可能有误,请以原文为准:Closed-Loop Physics-Based Manipulation Evaluation at a Commodity Price: What 4.4 L4 GPU hours at $3.50 Buys You(arXiv 2611.07441)

核心内容 ​

针对接触密集型机器人操作评估成本高昂的问题,研究团队发布了一套基于无服务器商用GPU的可复现评估工具。该工具在200次rollout规模下,单次复现耗时仅38分钟,消耗4.4个L4 GPU小时,成本约3.5美元,显著降低了物理仿真策略验证门槛,使中小团队也能进行统计学规模的闭环评估。

利用该工具进行的对照实验显示,在稀疏数据条件下,三个独立训练的动作分块(Action Chunking)检查点在780次rollout中成功率为零;相比之下,单步策略成功率介于2.6%至11.8%之间。通过逐步重规划控制变量法,确认失败源于多步预测的训练公式缺陷而非开环执行问题,为具身算法选型提供了关键负样本参考。

关键看点 ​

  • 评测成本:基于L4 GPU的无服务器架构,单次200-rollout复现成本3.5美元,耗时38分钟
  • 策略对比:稀疏数据下动作分块策略780次测试零成功,单步策略成功率2.6%~11.8%
  • 失效归因:逐步重规划实验证实故障源于多步预测训练公式,非开环执行机制问题
  • 开源资产:提供可复现的评估封装工具,支持在商用GPU上进行统计学规模物理仿真

工程与产业观察 ​

该工作以极低成本量化了动作分块策略在稀疏数据场景下的失效边界,警示工程师避免盲目套用热门范式。其提供的标准化评测Harness与负样本结论,为具身智能算法训练中的数据效率研究与仿真验证流程提供了高性价比的工程基准与避坑指南。

英文摘要 (Abstract) ​

Running evaluations of contact-rich robotic manipulation at statistical scale has largely been possible only at well-resourced labs, highlighting the need for affordable and efficient options for broader audiences. We present a reproducible evaluation harness wrapper based on a previous study at 200-rollout scale on serverless commodity GPUs, at a cost of approximately $3.50 USD and 38 minutes of wall-clock per replication (4.4 L4-GPU-hours across eight parallel shards). Using the harness, we report a partially controlled negative result on action chunking with sparse data: three independently trained chunked checkpoints produce zero genuine successes across 780 rollouts, while three single-step checkpoints range from 2.6% to 11.8% success rate. A per-step-replanning control partially isolates chunk prediction, rather than open-loop chunk execution as the mechanism, with the failure localized to the training formulation and multi-step prediction. This work further democratizes physics-based policy development and evaluation through our reusable harness.

作者:Rohun Kshirsagar, Diana Rypkema

基于本页技术方向关联的整机参数、核心零部件、实测报告与同赛道选型横评。

来源

本页事实对应的公开页面,以原文为准。

  1. arXiv 2611.07441:Closed-Loop Physics-Based Manipulation Evaluation at a Commodity Price: What 4.4 L4 GPU hours at $3.50 Buys You访问 2026-11-06