SQAM: A Scalar Approximation for Efficient Reinforcement Learning Fine-Tuning of Flow-Matching Robot Policies
机器人学会了用一个数字代替整个矩阵
Researchers propose SQAM (Q-Learning with Scalar Adjoint Matching) to address high computational costs in fine-tuning flow-matching robot policies. By observing that the Jacobian matrix is nearly diagonal, they approximate it with a scalar, enabling closed-form solutions for gradient propagation. Combined with value penalties to constrain Q-function estimates at generated actions, SQAM achieves higher success rates on complex navigation and manipulation tasks compared to baselines like TRQAM and EXPO-FT, validated on both simulation benchmarks and real dual-arm robots.
Source: Chinese robotics — Research 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 success rate
62%
Reported trials
View original evidence
| 任务领域 | 最强基线成功率 | SQAM成功率 | 提升幅度 | |---|---|---|---| | antmaze-giant | 41% | **62%** | 21个百分点 | | humanoidmaze-large | 36% | **54%** | 18个百分点 | | cube-triple | 50% | **72%** | 22个百分点 | | cube-quadruple | 19% | **54%** | 35个百分点 | 这四个领域恰恰是整个基准里最难啃的四个,数据集最大（最多到一亿条转移样本）、任务最复杂、动作空间维度最高。
Open source S18Reported success rate
54%
Reported trials
View original evidence
| 任务领域 | 最强基线成功率 | SQAM成功率 | 提升幅度 | |---|---|---|---| | antmaze-giant | 41% | **62%** | 21个百分点 | | humanoidmaze-large | 36% | **54%** | 18个百分点 | | cube-triple | 50% | **72%** | 22个百分点 | | cube-quadruple | 19% | **54%** | 35个百分点 | 这四个领域恰恰是整个基准里最难啃的四个,数据集最大（最多到一亿条转移样本）、任务最复杂、动作空间维度最高。
Open source S18Reported success rate
72%
Reported trials
View original evidence
| 任务领域 | 最强基线成功率 | SQAM成功率 | 提升幅度 | |---|---|---|---| | antmaze-giant | 41% | **62%** | 21个百分点 | | humanoidmaze-large | 36% | **54%** | 18个百分点 | | cube-triple | 50% | **72%** | 22个百分点 | | cube-quadruple | 19% | **54%** | 35个百分点 | 这四个领域恰恰是整个基准里最难啃的四个,数据集最大（最多到一亿条转移样本）、任务最复杂、动作空间维度最高。
Open source S18training_steps_offline
15,000
View original evidence
每个任务先用人类示范做监督微调，再用SQAM跑15000步离线强化学习，接着继续跑10000步在线强化学习,也就是让机器人边跟真实世界互动边继续学习。
Open source S21training_steps_online
10,000
View original evidence
每个任务先用人类示范做监督微调，再用SQAM跑15000步离线强化学习，接着继续跑10000步在线强化学习,也就是让机器人边跟真实世界互动边继续学习。
Open source S21
- paper: not_reported
What remains unknown
Not established in the collected evidence: Environment, 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.
真机实验:走出仿真房间 光在仿真器里好用还不够，研究者把SQAM搬到了真实的双臂机器人上，用的是一台配备两只20自由度灵巧手的人形机器人,底座是RB-Y1，手是Wuji Hand 2，脑子是一个叫RLDX-1的视觉语言动作预训练模型。
Open source S19
在整体五十个任务上，SQAM达到了75%的平均成功率，而之前最强的基线方法TRQAM是68%。
Open source S17
| 任务领域 | 最强基线成功率 | SQAM成功率 | 提升幅度 | |---|---|---|---| | antmaze-giant | 41% | **62%** | 21个百分点 | | humanoidmaze-large | 36% | **54%** | 18个百分点 | | cube-triple | 50% | **72%** | 22个百分点 | | cube-quadruple | 19% | **54%** | 35个百分点 | 这四个领域恰恰是整个基准里最难啃的四个,数据集最大（最多到一亿条转移样本）、任务最复杂、动作空间维度最高。
Open source S18
每个任务先用人类示范做监督微调，再用SQAM跑15000步离线强化学习，接着继续跑10000步在线强化学习,也就是让机器人边跟真实世界互动边继续学习。
Open source S21
Source：Chinese robotics — Research news discovery · msn.cn