发表机构
University of Bonn; Lamarr Institute for Machine Learning and Artificial Intelligence（波恩大学; 拉马尔机器学习和人工智能研究所）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
提出S$^3$方法，将冻结的文本到运动模型与仿真验证结合，无需训练即可提升人形机器人运动生成的可执行性，在HumanML3D上显著提高直立执行率，并验证了硬件可行性。
AI 中文摘要
文本到运动模型能生成看似合理的人体运动，但未对机器人的动力学进行建模；全身跟踪控制器能可靠地执行机器人参考轨迹，但无法重新规划不可行的参考轨迹。近期的语言到人形机器人系统通过训练来弥合这一差距。我们通过将部署控制器本身置于循环中，衡量在不进行任何训练的情况下能弥合多大差距。采样-仿真-筛选（S$^3$）从冻结的文本到运动模型中为每个提示词抽取$N$个运动，通过方向匹配逆运动学将每个运动重定向到Unitree G1，在完整刚体动力学下使用预训练的SONIC跟踪策略对所有运动进行滚动仿真，并保留策略执行效果最佳的候选。由于验证器是确定性仿真器本身，S$^3$通过构造达到了任意$N$上限；我们衡量的是该上限位于何处以及哪些因素未达到该上限。在200个分层HumanML3D测试提示词上，$N=8$时，直立执行率从83.5%提升至89.5%，硬件门控通过数从33提升至85；在完整测试集（4,184个提示词）上，直立执行率从80.5%提升至89.5%。一个能很好预测跌倒的动力学验证器（AUROC 0.90）仅恢复了这一增益的四分之一：对提示词自身的候选进行排序比分类总体更难。选择无法修复的一类问题是降低骨盆的提示词，而一个在重定向机器人数据上训练的生成器能够执行这类提示词。我们进一步使用标准文本-运动评估器对执行运动的语义保真度进行评分，并设置真实动作捕捉对照组以将损失归因于机器人投影，将重定向器与GMR进行消融比较（互补失败：两者的任意8上限提升至95.0%），并在真实G1上执行所有177个门控选择的片段：每个片段都以站立完成，硬件跟踪误差与仿真匹配（$r=0.94$）。
英文摘要
Text-to-motion models generate plausible human motion but do not model a robot's dynamics; whole-body tracking controllers execute robot references reliably but cannot replan an infeasible one. Recent language-to-humanoid systems bridge this gap by training. We measure how much of the gap closes with no training at all, by putting the deployment controller itself in the loop. Sample-simulate-select (S$^3$) draws $N$ motions per prompt from a frozen text-to-motion model, retargets each to a Unitree G1 by direction-matching inverse kinematics, rolls all of them out under full rigid-body dynamics with the pretrained SONIC tracking policy, and keeps the candidate the policy executed best. Because the verifier is the deterministic simulator itself, S$^3$ attains the any-of-$N$ ceiling by construction; what we measure is where that ceiling lies and what falls short of it. On 200 stratified HumanML3D test prompts with $N=8$, upright execution rises from 83.5% to 89.5% and hardware-gate passes from 33 to 85; on the complete test split (4,184 prompts) it rises from 80.5% to 89.5%. A kinematic verifier that predicts falls well (AUROC 0.90) recovers only a quarter of this gain: ranking a prompt's own candidates is harder than classifying the population. What selection cannot fix is one class, prompts that lower the pelvis, which a generator trained on retargeted robot data does execute. We further score the semantic fidelity of the executed motion with the standard text-motion evaluator, with a real-mocap control that attributes the loss to the robot projection, ablate the retargeter against GMR (complementary failures: the any-of-8 ceiling rises to 95.0% over both), and execute all 177 gate-selected clips on the real G1: every one completes standing, with hardware tracking error matching simulation ($r=0.94$).
Comments8 pages, 9 figures, 5 tables
