Add concise paper-aligned metric descriptions to RoboValue dataset cards

README.md

@@ -45,7 +45,7 @@ RoboValue provides expert demonstrations for **robotic value models**, covering
 All four training subsets are publicly available. Two **small-scale public diagnostic test sets** are organized separately under `evaluation/`, with **30 real-world trajectories and 30 simulation trajectories**. Each domain has four tasks and 90 synchronized camera videos. Together the test subsets contain **60 trajectories, 44,647 time steps, and 180 videos**. [The paper](https://arxiv.org/abs/2610.26663) and current leaderboard report **Zero-Shot and One-Shot evaluations**.
 
 <p align="center">
-  <a href="#overview">Overview</a> · <a href="#simulation">Simulation</a> · <a href="#real-world">Real-World Training</a> · <a href="#public-evaluation">Public Test Sets</a> · <a href="#repository-layout">Repository</a> · <a href="#download-and-reading">Download &amp; Read</a>
+  <a href="#overview">Overview</a> · <a href="#simulation">Simulation</a> · <a href="#real-world">Real-World Training</a> · <a href="#public-evaluation">Public Test Sets</a> · <a href="#metric-summary">Metrics</a> · <a href="#repository-layout">Repository</a> · <a href="#download-and-reading">Download &amp; Read</a>
 </p>
 
 ## 📊 1. Dataset Overview <a name="overview"></a>
@@ -93,6 +93,28 @@ See the [real-world metric coverage and trajectory index](evaluation/real-world/
 
 Paper-linked annotation lookups now provide the printed S1-Sn definitions and page references for all 35 tasks, plus mapped copies of all existing temporal labels. See [the reading examples](TASK_ALIGNMENT.md#paper-linked-annotations).
 
+<a id="metric-summary"></a>
+
+### Metric summary
+
+These one-sentence definitions follow the paper’s §4.3 (pp. 9–12). **↑ Higher is better; FPL ↓ lower is better.**
+
+| Metric | Brief definition |
+| --- | --- |
+| **SA · Success Accuracy ↑** | Measures the fraction of same-task successful/unsuccessful episode pairs in which the successful episode has the higher average terminal score. |
+| **TGA-CT · Task Grounding Accuracy - Cross-Task ↑** | Measures how often the correct instruction gives a larger start-to-end value gain than every cross-task negative instruction for the same successful episode. |
+| **TGA-CF · Task Grounding Accuracy - Counterfactual ↑** | Measures how often the correct instruction gives a larger start-to-end value gain than every counterfactual instruction changing the object, action, placement, or constraint. |
+| **SIA · Subtask Identification Accuracy ↑** | Scores subtask recognition as the geometric mean of the independent judge’s probabilities for the correct subtask, based on the model’s free-form descriptions. |
+| **VOC · Value-Order Correlation ↑** | Measures the Spearman rank correlation between predicted values and normalized time along successful trajectories. |
+| **Cycle-VOC ↑** | Averages Spearman correlations with forward and reversed progress order when a successful trajectory is replayed forward then backward as one continuous sequence. |
+| **Memory-VOC (VOC-MEM) ↑** | Applies VOC’s Spearman score to successful long trajectories with recurring visual states, using the history available to each model interface. |
+| **FPL · Failure-Point Localization ↓** | Measures the absolute time error between the annotated error onset and the onset inferred from the largest detected value decline, normalized by trajectory duration. |
+| **TRR · Trajectory Recovery Reasoning ↑** | Checks matched trajectories for the expected stage trends: failure down, continued error non-increasing, recovery attempt up, and successful/failed recovery outcome up/down. |
+| **VS · Value Stability ↑** | Combines trend efficiency with non-flat time coverage to reward stable value updates while penalizing reversals and prolonged flat intervals. |
+| **CSVC · Cross-Solution Value Consistency ↑** | Measures the consistency of local value gains for semantically matched subtasks across valid execution solutions. |
+
+The paper’s **Memory-VOC** is labeled **VOC-MEM** in this dataset, with membership key `voc_mem`. Stored labels and keys retain these names. VOC and VOC-MEM annotations are sampling anchors, not calibrated per-frame value scores.
+
 ## 🟦 2. Simulation Training <a name="simulation"></a>
 
 ### 2.1 Overview

README_zh.md

@@ -16,7 +16,7 @@ RoboValue 提供面向**机器人价值模型**的专家示例，覆盖 **15 个
 本仓库的四个训练子集均已完整公开。另在独立的 `evaluation/` 目录组织两个**小规模公开诊断测试集**：**真机30条、仿真30条轨迹**，每个领域各含4个任务、90个同步相机视频。两个测试子集合计 **60条轨迹、44,647个时间步和180个视频**。[论文](https://arxiv.org/abs/2610.26663)与当前排行榜只报告 **Zero-Shot 和 One-Shot 评测结果**。
 
 <p align="center">
-  <a href="#overview">数据总览</a> · <a href="#simulation">仿真世界</a> · <a href="#real-world">真实世界训练数据</a> · <a href="#public-evaluation">公开测试集</a> · <a href="#repository-layout">仓库目录</a> · <a href="#download-and-reading">下载与读取</a>
+  <a href="#overview">数据总览</a> · <a href="#simulation">仿真世界</a> · <a href="#real-world">真实世界训练数据</a> · <a href="#public-evaluation">公开测试集</a> · <a href="#metric-summary">指标简述</a> · <a href="#repository-layout">仓库目录</a> · <a href="#download-and-reading">下载与读取</a>
 </p>
 
 ## 📊 1. 数据总览 <a name="overview"></a>
@@ -64,6 +64,28 @@ RoboValue 提供面向**机器人价值模型**的专家示例，覆盖 **15 个
 
 现提供全部35任务的论文S编号、定义及页码索引，并将全部已有时间标注对应到论文词条。详见[读取说明](TASK_ALIGNMENT_zh.md#paper-linked-annotations)。
 
+<a id="metric-summary"></a>
+
+### 指标简述
+
+以下每项一句话，对应论文第4.3节（第9–12页）。**↑ 越高越好；FPL ↓ 越低越好。**
+
+| 指标 | 简要说明 |
+| --- | --- |
+| **SA · Success Accuracy ↑** | 同任务成功与失败轨迹对中，成功轨迹末段平均评分高于失败轨迹的比例。 |
+| **TGA-CT · Task Grounding Accuracy - Cross-Task ↑** | 固定成功轨迹，正确指令的首尾价值增益高于所有跨任务负指令的比例。 |
+| **TGA-CF · Task Grounding Accuracy - Counterfactual ↑** | 正确指令的首尾价值增益高于所有改变物体、动作、放置或约束的反事实指令的比例。 |
+| **SIA · Subtask Identification Accuracy ↑** | 根据模型自由生成的子任务描述，取独立评审赋予真实子任务概率的几何平均。 |
+| **VOC · Value-Order Correlation ↑** | 成功轨迹中，预测价值与归一化时间之间的Spearman秩相关。 |
+| **Cycle-VOC ↑** | 成功轨迹先正放再倒放，平均两段价值与对应正、逆进度顺序的Spearman秩相关。 |
+| **Memory-VOC（VOC-MEM） ↑** | 使用模型接口可用的执行历史，在视觉状态重复的成功长轨迹上计算VOC。 |
+| **FPL · Failure-Point Localization ↓** | 最大检测价值下降所定位的错误起点与标注起点的绝对时间差，按轨迹时长归一化。 |
+| **TRR · Trajectory Recovery Reasoning ↑** | 检查错误时下降、延续错误时不升、恢复尝试时上升，以及恢复成功升、失败降的阶段趋势。 |
+| **VS · Value Stability ↑** | 结合趋势效率比与非平坦时间覆盖，衡量少反转且持续提供价值更新的稳定性。 |
+| **CSVC · Cross-Solution Value Consistency ↑** | 比较不同有效执行方案中，同一语义子任务的局部价值增益一致性。 |
+
+论文的 **Memory-VOC** 在本数据中标为 **VOC-MEM**，指标键为 `voc_mem`，存储名称保持一致。VOC／VOC-MEM 标注是采样锚点，并非逐帧标定价值分数。
+
 ## 🟦 2. 仿真训练数据 <a name="simulation"></a>
 
 ### 2.1 总览

evaluation/real-world/test-30/README.md

@@ -4,7 +4,7 @@
 
 This **in-domain diagnostic test subset** contains **30 complete real-world trajectories across four tasks**, including standard demonstrations, alternative successful solutions, errors, recoveries, and counting histories. It provides visual inputs, English instructions, and separate evaluation annotations for **11 metric entry points**. Its purpose is to make metric behavior inspectable and support small-scale evaluation; its task and solution coverage is smaller than the complete RoboValue leaderboard test set.
 
-The subset is stored under `evaluation/real-world/test-30/`, separately from the repository's training demonstrations in `data/`. The four tasks belong to existing real-world task categories; evaluation coverage is defined by the metric cohorts below. Native metadata retains `train: 0:30` for LeRobot container compatibility; the release's intended use and HF trajectory-catalog split are **test**.
+The subset is stored under `evaluation/real-world/test-30/`, separately from the repository's training demonstrations in `training-data/`. The four tasks belong to existing real-world task categories; evaluation coverage is defined by the metric cohorts below. Native metadata retains `train: 0:30` for LeRobot container compatibility; the release's intended use and HF trajectory-catalog split are **test**.
 
 ## Tasks and trajectories
 
@@ -25,19 +25,21 @@ Local `task_index` values 0–3 and `episode_index` values 0–29 belong to this
 
 [`annotations/metric_membership.json`](annotations/metric_membership.json) is the **authoritative metric membership**, even when other annotation fields exist on an episode. The same episode can participate in several metrics, so the counts below must not be added to obtain the dataset size.
 
-| Metric / code key | What it evaluates | Eligible trajectories | Episodes / groups | Supporting annotations |
+Definitions follow the paper’s §4.3 (pp. 9–12): **higher is better except FPL (lower is better)**. The paper’s Memory-VOC uses the dataset label VOC-MEM and key `voc_mem`. See the [complete metric summary](../../../README.md#metric-summary).
+
+| Metric / code key | Brief definition | Eligible trajectories | Episodes / groups | Supporting annotations |
 |---|---|---|---|---|
-| TGA-CT / `tga_easy` | Distinguishing the correct task instruction from other tasks | Desk and flower ED | 4 episodes | `tasks.json`, `task_candidates.json` (14 candidate instructions) |
-| TGA-CF / `tga_hard` | Distinguishing the task goal from counterfactual instructions | Desk and flower ED | 4 episodes | `tasks.json` (4 counterfactual instructions per regular task) |
-| VOC / `voc` | Value ordering along forward task progress | Desk and flower ED | 4 episodes | Subtask intervals and sampling anchors in `episodes.jsonl` |
-| Cycle-VOC / `cycle_voc` | Value ordering along forward and reverse task progress | Desk and flower ED | 4 episodes | Complete trajectories, subtask intervals, and sampling anchors in `episodes.jsonl` |
-| VS / `vs` | Stability of the forward value curve | Desk and flower ED | 4 episodes | The same forward trajectories and sampling basis used by Cycle-VOC |
-| SA / `sa` | Distinguishing terminal task success from failure | All desk and flower episodes | 24 episodes: 16 successful, 8 unsuccessful | `success` in `episodes.jsonl` |
-| SIA / `sia` | Identifying the semantic subtask being performed | Desk and flower ED/MS | 12 episodes | Subtask descriptions in `tasks.json` and local intervals in `episodes.jsonl` |
-| CSVC / `csvc` | Consistency of semantic value increments across successful solutions | Desk and flower ED/MS | 12 episodes / 10 solution groups | `csvc_solutions.json`: groups, semantic/object order, and subtask intervals |
-| TRR / `trr` | Distinguishing effective recovery, ineffective recovery, and error continuation | Desk and flower ER/IR/EC | 12 episodes / 4 three-role groups | `trr_groups.json` and failure/recovery stage intervals in `episodes.jsonl` |
-| FPL / `fpl` | Locating the onset of the target error | Selected desk and flower ER/EC | 8 episodes: 4 ER, 4 EC | `fpl_annotations.json`: target subtask, failure interval, and `failure_onset_frame` |
-| VOC-MEM / `voc_mem` | Value ordering for repeated-action counting using complete history | Drum and button MEM | 6 episodes, all successful | Counting subtask intervals and sampling anchors in `episodes.jsonl` |
+| TGA-CT / `tga_easy` | Measures how often the correct instruction gives a larger start-to-end value gain than every cross-task negative instruction for the same successful episode. | Desk and flower ED | 4 episodes | `tasks.json`, `task_candidates.json` (14 candidate instructions) |
+| TGA-CF / `tga_hard` | Measures how often the correct instruction gives a larger start-to-end value gain than every counterfactual instruction changing the object, action, placement, or constraint. | Desk and flower ED | 4 episodes | `tasks.json` (4 counterfactual instructions per regular task) |
+| VOC / `voc` | Measures the Spearman rank correlation between predicted values and normalized time along successful trajectories. | Desk and flower ED | 4 episodes | Subtask intervals and sampling anchors in `episodes.jsonl` |
+| Cycle-VOC / `cycle_voc` | Averages Spearman correlations with forward and reversed progress order when a successful trajectory is replayed forward then backward as one continuous sequence. | Desk and flower ED | 4 episodes | Complete trajectories, subtask intervals, and sampling anchors in `episodes.jsonl` |
+| VS / `vs` | Combines trend efficiency with non-flat time coverage to reward stable value updates while penalizing reversals and prolonged flat intervals. | Desk and flower ED | 4 episodes | The same forward trajectories and sampling basis used by Cycle-VOC |
+| SA / `sa` | Measures the fraction of same-task successful/unsuccessful episode pairs in which the successful episode has the higher average terminal score. | All desk and flower episodes | 24 episodes: 16 successful, 8 unsuccessful | `success` in `episodes.jsonl` |
+| SIA / `sia` | Scores subtask recognition as the geometric mean of the independent judge’s probabilities for the correct subtask, based on the model’s free-form descriptions. | Desk and flower ED/MS | 12 episodes | Subtask descriptions in `tasks.json` and local intervals in `episodes.jsonl` |
+| CSVC / `csvc` | Measures the consistency of local value gains for semantically matched subtasks across valid execution solutions. | Desk and flower ED/MS | 12 episodes / 10 solution groups | `csvc_solutions.json`: groups, semantic/object order, and subtask intervals |
+| TRR / `trr` | Checks matched trajectories for the expected stage trends: failure down, continued error non-increasing, recovery attempt up, and successful/failed recovery outcome up/down. | Desk and flower ER/IR/EC | 12 episodes / 4 three-role groups | `trr_groups.json` and failure/recovery stage intervals in `episodes.jsonl` |
+| FPL / `fpl` | Measures the absolute time error between the annotated error onset and the onset inferred from the largest detected value decline, normalized by trajectory duration. | Selected desk and flower ER/EC | 8 episodes: 4 ER, 4 EC | `fpl_annotations.json`: target subtask, failure interval, and `failure_onset_frame` |
+| VOC-MEM / `voc_mem` | Applies VOC’s Spearman score to successful long trajectories with recurring visual states, using the history available to each model interface. | Drum and button MEM | 6 episodes, all successful | Counting subtask intervals and sampling anchors in `episodes.jsonl` |
 
 All four ED episodes are `0, 1, 12, 14`. Counting episodes are `24–29`. For a per-episode list, inspect [`episode_catalog.parquet`](episode_catalog.parquet), whose `metrics` column is derived from the membership file.
 

evaluation/real-world/test-30/README_zh.md

@@ -4,7 +4,7 @@
 
 这是一个**真实世界 In-Domain 诊断测试子集**，包含 **4 个任务、30 条完整轨迹**，涵盖标准示范、多种成功方案、错误、恢复和完整计数历史。数据提供图像、英语任务指令及独立评测标注，支持 **11 个指标入口**，便于检查指标行为和开展小规模评测。它的任务与方案覆盖范围小于完整 RoboValue 排行榜测试集。
 
-本子集位于 `evaluation/real-world/test-30/`，与仓库 `data/` 下的训练示范分区存放，覆盖已有的4个真机任务类别；评测范围以下方各指标的正式轨迹集合为准。为兼容 LeRobot 容器，原生元数据保留 `train: 0:30`；本次发布的用途及 HF 逐轨迹目录划分均为 **test（测试）**。
+本子集位于 `evaluation/real-world/test-30/`，与仓库 `training-data/` 下的训练示范分区存放，覆盖已有的4个真机任务类别；评测范围以下方各指标的正式轨迹集合为准。为兼容 LeRobot 容器，原生元数据保留 `train: 0:30`；本次发布的用途及 HF 逐轨迹目录划分均为 **test（测试）**。
 
 ## 任务与轨迹
 
@@ -25,19 +25,21 @@
 
 [`annotations/metric_membership.json`](annotations/metric_membership.json) 是**唯一权威的指标归属表**。即使某条轨迹还含有其他标注字段，也应按该文件筛选正式评测集合。同一轨迹可参加多个指标，不能把下表的轨迹数相加作为数据集总量。
 
-| 指标／代码键 | 评测用途 | 适用轨迹 | 轨迹数／组数 | 对应标注 |
+定义对应论文第4.3节（第9–12页）：**除FPL越低越好外，其余指标越高越好**。论文的Memory-VOC在数据中标为VOC-MEM，指标键为 `voc_mem`。完整指标见[首页简述](../../../README_zh.md#metric-summary)。
+
+| 指标／代码键 | 简要说明 | 适用轨迹 | 轨迹数／组数 | 对应标注 |
 |---|---|---|---|---|
-| TGA-CT／`tga_easy` | 从不同任务的指令中辨别正确任务目标 | 书桌、花艺 ED | 4条 | `tasks.json`、`task_candidates.json`（14个候选指令） |
-| TGA-CF／`tga_hard` | 区分原任务目标与反事实任务指令 | 书桌、花艺 ED | 4条 | `tasks.json`（每个常规任务4条反事实指令） |
-| VOC／`voc` | 检查正向任务进展的价值顺序 | 书桌、花艺 ED | 4条 | `episodes.jsonl` 中的子任务区间和采样锚点 |
-| Cycle-VOC／`cycle_voc` | 检查正向及反向播放下的进展价值顺序 | 书桌、花艺 ED | 4条 | 完整轨迹、子任务区间、采样锚点 |
-| VS／`vs` | 检查正向价值曲线的稳定性 | 书桌、花艺 ED | 4条 | 与 Cycle-VOC 共享的正向轨迹和采样依据 |
-| SA／`sa` | 区分最终任务成功与失败 | 书桌、花艺全部轨迹 | 24条：16成功、8失败 | `episodes.jsonl` 中的 `success` |
-| SIA／`sia` | 识别当前执行的语义子任务 | 书桌、花艺 ED/MS | 12条 | `tasks.json` 的子任务描述及 `episodes.jsonl` 的局部区间 |
-| CSVC／`csvc` | 检查不同成功方案中的语义价值增量一致性 | 书桌、花艺 ED/MS | 12条／10个方案组 | `csvc_solutions.json` 中的分组、语义／物体顺序及子任务区间 |
-| TRR／`trr` | 区分有效恢复、无效恢复和错误延续 | 书桌、花艺 ER/IR/EC | 12条／4个三角色组 | `trr_groups.json` 及 `episodes.jsonl` 的失败／恢复阶段区间 |
-| FPL／`fpl` | 定位目标错误的发生起点 | 选定的书桌、花艺 ER/EC | 8条：4 ER、4 EC | `fpl_annotations.json` 中的目标子任务、错误区间、`failure_onset_frame` |
-| VOC-MEM／`voc_mem` | 利用完整历史判断重复动作的计数进展顺序 | 敲鼓、按钮 MEM | 6条，全部最终成功 | `episodes.jsonl` 中的计数子任务区间和采样锚点 |
+| TGA-CT／`tga_easy` | 固定成功轨迹，正确指令的首尾价值增益高于所有跨任务负指令的比例。 | 书桌、花艺 ED | 4条 | `tasks.json`、`task_candidates.json`（14个候选指令） |
+| TGA-CF／`tga_hard` | 正确指令的首尾价值增益高于所有改变物体、动作、放置或约束的反事实指令的比例。 | 书桌、花艺 ED | 4条 | `tasks.json`（每个常规任务4条反事实指令） |
+| VOC／`voc` | 成功轨迹中，预测价值与归一化时间之间的Spearman秩相关。 | 书桌、花艺 ED | 4条 | `episodes.jsonl` 中的子任务区间和采样锚点 |
+| Cycle-VOC／`cycle_voc` | 成功轨迹先正放再倒放，平均两段价值与对应正、逆进度顺序的Spearman秩相关。 | 书桌、花艺 ED | 4条 | 完整轨迹、子任务区间、采样锚点 |
+| VS／`vs` | 结合趋势效率比与非平坦时间覆盖，衡量少反转且持续提供价值更新的稳定性。 | 书桌、花艺 ED | 4条 | 与 Cycle-VOC 共享的正向轨迹和采样依据 |
+| SA／`sa` | 同任务成功与失败轨迹对中，成功轨迹末段平均评分高于失败轨迹的比例。 | 书桌、花艺全部轨迹 | 24条：16成功、8失败 | `episodes.jsonl` 中的 `success` |
+| SIA／`sia` | 根据模型自由生成的子任务描述，取��立评审赋予真实子任务概率的几何平均。 | 书桌、花艺 ED/MS | 12条 | `tasks.json` 的子任务描述及 `episodes.jsonl` 的局部区间 |
+| CSVC／`csvc` | 比较不同有效执行方案中，同一语义子任务的局部价值增益一致性。 | 书桌、花艺 ED/MS | 12条／10个方案组 | `csvc_solutions.json` 中的分组、语义／物体顺序及子任务区间 |
+| TRR／`trr` | 检查错误时下降、延续错误时不升、恢复尝试时上升，以及恢复成功升、失败降的阶段趋势。 | 书桌、花艺 ER/IR/EC | 12条／4个三角色组 | `trr_groups.json` 及 `episodes.jsonl` 的失败／恢复阶段区间 |
+| FPL／`fpl` | 最大检测价值下降所定位的错误起点与标注起点的绝对时间差，按轨迹时长归一化。 | 选定的书桌、花艺 ER/EC | 8条：4 ER、4 EC | `fpl_annotations.json` 中的目标子任务、错误区间、`failure_onset_frame` |
+| VOC-MEM／`voc_mem` | 使用模型接口可用的执行历史，在视觉状态重复的成功长轨迹上计算VOC。 | 敲鼓、按钮 MEM | 6条，全部最终成功 | `episodes.jsonl` 中的计数子任务区间和采样锚点 |
 
 4条 ED 的编号为 `0、1、12、14`；计数轨迹的编号为 `24–29`。每条轨迹的适用指标也可在 [`episode_catalog.parquet`](episode_catalog.parquet) 的 `metrics` 列查看，该列由权威归属表反向生成。
 

evaluation/real-world/test-30/checksums.sha256

@@ -1,5 +1,5 @@
-623c786a5cc5ab4d4d9cb985c181d1c7c2d5bb2f27a40aba517953477c718ba8  README.md
-e3a0ecf454a4113e0748d8983db04484fc3c8e530ae65eb6dcdcf689eb2bc66f  README_zh.md
+37b8c2a4d4b272649117f73615f5583a214c866613d8468ed87dcd98f59bc599  README.md
+e74f1c3c3a3c0f5ac53eeb1cc8f680c99b70058eed8453996797315232ec238e  README_zh.md
 a377ffb50a1adf5354a8d65e9a90175b73cacb4a9800593cf1eaf89885f2f1df  TRAJECTORIES.md
 3cea0680a8709e160e87ec3c01f6b355a3e3956ee68c6b9ef5e628269c792e5a  annotations/csvc_solutions.json
 a3be4d8110975a3a3214b4446a5d31cf41b99a91dc104d5a8f45b3129635806b  annotations/episodes.jsonl

evaluation/simulation/test-30/README.md

@@ -4,7 +4,7 @@
 
 This **In-Domain diagnostic test subset** contains **30 complete simulation trajectories across four tasks**. It provides RGB video inputs, English instructions, and separate annotations for **10 metric entry points**. Standard demonstrations, alternative orders, error/recovery branches, and complete memory histories are retained. Its task and solution coverage is smaller than the complete RoboValue leaderboard test set.
 
-The subset is stored under `evaluation/simulation/test-30/`, separately from training demonstrations in `data/`. Local episode and task indices belong to this subset. Native metadata retains `train: 0:30` for LeRobot container compatibility; the intended use and HF trajectory-catalog split are **test**.
+The subset is stored under `evaluation/simulation/test-30/`, separately from training demonstrations in `training-data/`. Local episode and task indices belong to this subset. Native metadata retains `train: 0:30` for LeRobot container compatibility; the intended use and HF trajectory-catalog split are **test**.
 
 ## Tasks and trajectories
 
@@ -27,18 +27,20 @@ Trajectory types are **ED4 / MS8 / ER4 / IR4 / EC4 / MEM6**. Stored terminal-suc
 
 [`annotations/metric_membership.json`](annotations/metric_membership.json) is the authoritative membership file. The same asset can participate in several metrics: **86 metric-member references correspond to 30 stored episodes**.
 
-| Metric / code key | What it evaluates | Eligible trajectories | Episodes / groups | Supporting annotations |
+Definitions follow the paper’s §4.3 (pp. 9–12): **higher is better except FPL (lower is better)**. The paper’s Memory-VOC uses the dataset label VOC-MEM and key `voc_mem`. See the [complete metric summary](../../../README.md#metric-summary).
+
+| Metric / code key | Brief definition | Eligible trajectories | Episodes / groups | Supporting annotations |
 | --- | --- | --- | --- | --- |
-| TGA-CT / `tga_easy` | Distinguishing the correct task from other task instructions | Pen-holder and mug ED | 4 | `tasks.json`, `task_candidates.json` (13 original candidates) |
-| TGA-CF / `tga_hard` | Distinguishing the original goal from counterfactual instructions | Pen-holder and mug ED | 4 | `tasks.json` (4 counterfactual instructions per regular task) |
-| VOC / `voc` | Value ordering along forward task progress | Pen-holder and mug ED | 4 | Original subtask intervals and sampling anchors in `episodes.jsonl` |
-| Cycle-VOC / `cycle_voc` | Value ordering along forward and reverse task progress | Pen-holder and mug ED | 4 | Complete episodes, original intervals, and sampling anchors |
-| VS / `vs` | Stability of the forward value curve | Pen-holder and mug ED | 4 | Same forward episodes and sampling basis used by Cycle-VOC |
-| SA / `sa` | Comparing terminal values under supplied success/failure labels | All pen-holder and mug trajectories | 24; 64 same-task pairs | Stored labels and evidence states in `episodes.jsonl`, original pairings in `sa_pairs.jsonl` |
-| SIA / `sia` | Identifying the chronological semantic subtask being performed | Pen-holder and mug ED/MS | 12 | Original subtask descriptions and intervals |
-| TRR / `trr` | Distinguishing effective recovery, ineffective recovery, and error continuation | Pen-holder and mug ER/IR/EC | 12 / 4 three-role groups | `trr_groups.json`, original failure/recovery stages in `episodes.jsonl` |
-| CSVC / `csvc` | Consistency of semantic value increments across execution orders | Pen-holder and mug ED/MS | 12 / 6 order groups | `csvc_solutions.json`, independent `semantic_alignment.json` |
-| VOC-MEM / `voc_mem` | History-dependent value ordering in repeated or sequential actions | Press-by-number and swap-blocks MEM | 6 | Complete videos and original `voc_mem` sampling anchors |
+| TGA-CT / `tga_easy` | Measures how often the correct instruction gives a larger start-to-end value gain than every cross-task negative instruction for the same successful episode. | Pen-holder and mug ED | 4 | `tasks.json`, `task_candidates.json` (13 original candidates) |
+| TGA-CF / `tga_hard` | Measures how often the correct instruction gives a larger start-to-end value gain than every counterfactual instruction changing the object, action, placement, or constraint. | Pen-holder and mug ED | 4 | `tasks.json` (4 counterfactual instructions per regular task) |
+| VOC / `voc` | Measures the Spearman rank correlation between predicted values and normalized time along successful trajectories. | Pen-holder and mug ED | 4 | Original subtask intervals and sampling anchors in `episodes.jsonl` |
+| Cycle-VOC / `cycle_voc` | Averages Spearman correlations with forward and reversed progress order when a successful trajectory is replayed forward then backward as one continuous sequence. | Pen-holder and mug ED | 4 | Complete episodes, original intervals, and sampling anchors |
+| VS / `vs` | Combines trend efficiency with non-flat time coverage to reward stable value updates while penalizing reversals and prolonged flat intervals. | Pen-holder and mug ED | 4 | Same forward episodes and sampling basis used by Cycle-VOC |
+| SA / `sa` | Measures the fraction of same-task successful/unsuccessful episode pairs in which the successful episode has the higher average terminal score. | All pen-holder and mug trajectories | 24; 64 same-task pairs | Stored labels and evidence states in `episodes.jsonl`, original pairings in `sa_pairs.jsonl` |
+| SIA / `sia` | Scores subtask recognition as the geometric mean of the independent judge’s probabilities for the correct subtask, based on the model’s free-form descriptions. | Pen-holder and mug ED/MS | 12 | Original subtask descriptions and intervals |
+| TRR / `trr` | Checks matched trajectories for the expected stage trends: failure down, continued error non-increasing, recovery attempt up, and successful/failed recovery outcome up/down. | Pen-holder and mug ER/IR/EC | 12 / 4 three-role groups | `trr_groups.json`, original failure/recovery stages in `episodes.jsonl` |
+| CSVC / `csvc` | Measures the consistency of local value gains for semantically matched subtasks across valid execution solutions. | Pen-holder and mug ED/MS | 12 / 6 order groups | `csvc_solutions.json`, independent `semantic_alignment.json` |
+| VOC-MEM / `voc_mem` | Applies VOC’s Spearman score to successful long trajectories with recurring visual states, using the history available to each model interface. | Press-by-number and swap-blocks MEM | 6 | Complete videos and original `voc_mem` sampling anchors |
 
 The four ED episodes are `0, 1, 12, 13`; MEM episodes are `24–29`. The catalog's `metrics` column is generated from this membership file. Counts across metrics overlap and must not be summed to obtain the number of trajectories.
 

evaluation/simulation/test-30/README_zh.md

@@ -4,7 +4,7 @@
 
 这是一个 **In-Domain 诊断测试子集**，包含 **4个任务、30条完整仿真轨迹**，提供RGB视频输入、英文指令，以及独立存放的 **10个指标入口**所需标注。标准示范、替代执行顺序、错误／恢复分支和完整记忆历史均保留。它的任务与方案覆盖范围小于完整 RoboValue 排行榜测试集。
 
-本子集位于 `evaluation/simulation/test-30/`，与 `data/` 下的训练示范分开。任务和轨迹编号仅在本子集内有效。原生元数据保留 `train: 0:30` 以兼容 LeRobot 容器；公开用途及HF逐轨迹目录划分为 **test（测试）**。
+本子集位于 `evaluation/simulation/test-30/`，与 `training-data/` 下的训练示范分开。任务和轨迹编号仅在本子集内有效。原生元数据保留 `train: 0:30` 以兼容 LeRobot 容器；公开用途及HF逐轨迹目录划分为 **test（测试）**。
 
 ## 任务与轨迹
 
@@ -27,18 +27,20 @@
 
 [`annotations/metric_membership.json`](annotations/metric_membership.json) 是正式指标归属表。同一条轨迹可参与多个指标，**86次指标成员引用对应30条实际轨迹**。
 
-| 指标／代码键 | 评测用途 | 适用轨迹 | 轨迹数／组数 | 对应标注 |
+定义对应论文第4.3节（第9–12页）：**除FPL越低越好外，其余指标越高越好**。论文的Memory-VOC在数据中标为VOC-MEM，指标键为 `voc_mem`。完整指标见[首页简述](../../../README_zh.md#metric-summary)。
+
+| 指标／代码键 | 简要说明 | 适用轨迹 | 轨迹数／组数 | 对应标注 |
 | --- | --- | --- | --- | --- |
-| TGA-CT／`tga_easy` | 在不同任务指令中辨别正确目标 | 笔筒、马克杯ED | 4条 | `tasks.json`、`task_candidates.json`（原13个候选） |
-| TGA-CF／`tga_hard` | 区分原目标与反事实任务指令 | 笔筒、马克杯ED | 4条 | `tasks.json`（每个常规任务4条反事实指令） |
-| VOC／`voc` | 检查正向任务进展的价值顺序 | 笔筒、马克杯ED | 4条 | `episodes.jsonl` 原子任务区间及采样锚点 |
-| Cycle-VOC／`cycle_voc` | 检查正向及反向进展的价值顺序 | 笔筒、马克杯ED | 4条 | 完整轨迹、原区间及采样锚点 |
-| VS／`vs` | 检查正向价值曲线的稳定性 | 笔筒、马克杯ED | 4条 | 与Cycle-VOC共享的正向轨迹和采样依据 |
-| SA／`sa` | 在已有成功／失败标签下比较终局价值 | 笔筒、马克杯全部轨迹 | 24条；64个同任务配对 | `episodes.jsonl` 的原标签与证据状态、`sa_pairs.jsonl` 的原配对 |
-| SIA／`sia` | 识别按时间执行的语义子任务 | 笔筒、马克杯ED/MS | 12条 | 原子任务描述及区间 |
-| TRR／`trr` | 区分有效恢复、无效恢复和错误延续 | 笔筒、马克杯ER/IR/EC | 12条／4个三角色组 | `trr_groups.json`、`episodes.jsonl` 原失败／恢复阶段 |
-| CSVC／`csvc` | 检查不同顺序中的语义价值增量一致性 | 笔筒、马克杯ED/MS | 12条／6个顺序组 | `csvc_solutions.json` 与独立 `semantic_alignment.json` |
-| VOC-MEM／`voc_mem` | 重复或顺序动作中依赖历史的价值顺序 | 依数字按按钮、交换积木MEM | 6条 | 完整视频及原 `voc_mem` 采样锚点 |
+| TGA-CT／`tga_easy` | 固定成功轨迹，正确指令的首尾价值增益高于所有跨任务负指令的比例。 | 笔筒、马克杯ED | 4条 | `tasks.json`、`task_candidates.json`（原13个候选） |
+| TGA-CF／`tga_hard` | 正确指令的首尾价值增益高于所有改变物体、动作、放置或约束的反事实指令的比例。 | 笔筒、马克杯ED | 4条 | `tasks.json`（每个常规任务4条反事实指令） |
+| VOC／`voc` | 成功轨迹中，预测价值与归一化时间之间的Spearman秩相关。 | 笔筒、马克杯ED | 4条 | `episodes.jsonl` 原子任务区间及采样锚点 |
+| Cycle-VOC／`cycle_voc` | 成功轨迹先正放再倒放，平均两段价值与对应正、逆进度顺序的Spearman秩相关。 | 笔筒、马克杯ED | 4条 | 完整轨迹、原区间及采样锚点 |
+| VS／`vs` | 结合趋势效率比与非平坦时间覆盖，衡量少反转且持续提供价值更新的稳定性。 | 笔筒、马克杯ED | 4条 | 与Cycle-VOC共享的正向轨迹和采样依据 |
+| SA／`sa` | 同任务成功与失败轨迹对中，成功轨迹末段平均评分高于失败轨迹的比例。 | 笔筒、马克杯全部轨迹 | 24条；64个同任务配对 | `episodes.jsonl` 的原标签与证据状态、`sa_pairs.jsonl` 的原配对 |
+| SIA／`sia` | 根据模型自由生成的子任务描述，取独立评审赋予真实子任务概率的几何平均。 | 笔筒、马克杯ED/MS | 12条 | 原子任务描述及区间 |
+| TRR／`trr` | 检查错误时下降、延续错误时不升、恢复尝试时上升，以及恢复成功升、失败降的阶段趋势。 | 笔筒、马克杯ER/IR/EC | 12条／4个三角色组 | `trr_groups.json`、`episodes.jsonl` 原失败／恢复阶段 |
+| CSVC／`csvc` | 比较不同有效执行方案中，同一语义子任务的局部价值增益一致性。 | 笔筒、马克杯ED/MS | 12条／6个顺序组 | `csvc_solutions.json` 与独立 `semantic_alignment.json` |
+| VOC-MEM／`voc_mem` | 使用模型接口可用的执行历史，在视觉状态重复的成功长轨迹上计算VOC。 | 依数字按按钮、交换积木MEM | 6条 | 完整视频及原 `voc_mem` 采样锚点 |
 
 4条ED编号为 `0、1、12、13`；MEM编号为 `24–29`。轨迹目录的 `metrics` 列由该归属表生成。指标之间会复用轨迹，不能将指标轨迹数相加作为数据总量。
 

evaluation/simulation/test-30/checksums.sha256

@@ -1,5 +1,5 @@
-9a927d43573cde8cd2435f470754290da8e237548ac8f86c114034f820e2d9a2  README.md
-81539f9c30b21d0e7f7d83ca5a2b9611038a09f347cb891036525ae8953004bb  README_zh.md
+bab2327d2971a64a23de270652095a30fe047707028eaf01b59988b18330c22a  README.md
+9a402d9a2e4601523ef510311679b7b08955ee1a5b1c09d6aa794ce69be0b8e8  README_zh.md
 36ef94a87991421de07478d7d6667ff1755a93e45a8b173372b18c5ce9bc99ca  TRAJECTORIES.md
 c6609d781c66a153962e5c5c9412ba6289c4f0372d23d5a8c2f84842fa657130  annotations/annotation_provenance.json
 272dbff320bc7531dc0977a02eaa0a439e70b574394df28ff1420bec0ed8317b  annotations/csvc_solutions.json

subset_catalog.json

@@ -499,7 +499,7 @@
       "fps": 30,
       "camera_streams_per_episode": 3,
       "videos": 90,
-      "size_bytes": 382134761,
+      "size_bytes": 382136958,
       "duration_seconds": 916.6,
       "splits": {
         "train": "0:30"
@@ -658,7 +658,7 @@
       "primary_camera": "observation.images.front",
       "primary_video_column": "video_primary",
       "annotation_evidence_note": "ED/MS terminal labels and semantic alignment are user-attested; TRR/MEM terminal-label evidence remains pending independent verification.",
-      "size_bytes": 594651987,
+      "size_bytes": 594653966,
       "data_revision": "578593ef28ee7a2d861ebf38162c928291b0d4a9",
       "size_scope": "entire subset directory, including paper mappings, documentation and integrity files",
       "paper_subtasks": {
@@ -694,5 +694,26 @@
     "paper_definition_catalog": "task_alignment.json#paper_definition_catalog",
     "distinct_paper_definitions": 155,
     "paper_mapping_revision": "5b494e84028aa6212b060de2f6014a00d569e8cd"
+  },
+  "metric_documentation": {
+    "english": "README.md#metric-summary",
+    "chinese": "README_zh.md#metric-summary",
+    "paper_section": "4.3",
+    "paper_pages": [
+      9,
+      10,
+      11,
+      12
+    ],
+    "paper_metric_aliases": {
+      "Memory-VOC": {
+        "dataset_label": "VOC-MEM",
+        "membership_key": "voc_mem"
+      }
+    },
+    "score_direction": {
+      "default": "higher_is_better",
+      "FPL": "lower_is_better"
+    }
   }
 }