ConfAL-WM-Dataset
收藏资源简介:
ConfAL-WM 数据集是 ConfAL-WM(置信度引导的主动学习用于动作条件世界模型)项目的评估工件,包含在 RoboTwin2.0 候选池(约 18,000 个片段)上预计算的输出结果。该数据集主要用于支持世界模型的训练、主动学习策略评估以及置信度估计研究。数据集包含多个子集:(1) 50 任务预筛选包,每个任务一个代表性片段,包含预测帧、潜变量、动作、置信度分数、风险地图和 JSON 元数据;(2) EVAC-v2 训练推理包,包含 7298 个由 C3 均值风险采集选中的片段,每个片段有预测帧、潜变量、动作等;(3) 密集置信度/风险输出包,包含 7277 个选中片段的密集置信度地图和风险统计;(4) 基线选择推理结果包,包含其他采集基线(如 GVL、LRMS、PRM-judge、随机等)选中的片段;(5) 基线加权帧级评分数据包,包含帧级评分用于额外加权基线实验;(6) YOLO 标注包,包含机器人臂轨迹(夹爪)的伪标签,格式为 YOLO 检测格式。所有数据均经过路径清洗和匿名化处理,确保无用户或机器信息泄露。该数据集适用于机器人学习、主动学习、置信度估计、世界模型评估等任务。
The ConfAL-WM dataset is the evaluation artifact of the ConfAL-WM (Confidence-guided Active Learning for Action-Conditioned World Models) project, containing precomputed outputs on the RoboTwin2.0 candidate pool (approximately 18,000 episodes). It is primarily used to support world model training, active learning strategy evaluation, and confidence estimation research. The dataset includes multiple subsets: (1) 50-task prescreening package, one representative episode per task, containing predicted frames, latent variables, actions, confidence scores, risk maps, and JSON metadata; (2) EVAC-v2 training inference package, containing 7,298 episodes selected by C3 mean risk acquisition, each with predicted frames, latent variables, actions, etc.; (3) Dense confidence/risk output package, containing dense confidence maps and risk statistics for 7,277 selected episodes; (4) Baseline selection inference results package, containing episodes selected by other acquisition baselines (e.g., GVL, LRMS, PRM-judge, random); (5) Baseline weighted frame-level scoring data package, containing frame-level scores for additional weighted baseline experiments; (6) YOLO annotation package, containing pseudo-labels for robot arm trajectories (grippers) in YOLO detection format. All data have been path-cleaned and anonymized to ensure no leakage of user or machine information. The dataset is suitable for tasks such as robot learning, active learning, confidence estimation, and world model evaluation.
ConfAL-WM 数据集概述
ConfAL-WM: Confidence-Guided Active Learning for Action-Conditioned World Models(匿名投稿)的配套数据与评估产物,与项目主页“07 · Models & Data”部分对应。数据集提供预计算推理结果,避免对 RoboTwin2.0 候选池(约 18k 条轨迹)进行重复昂贵推理。
许可证与标签
- 许可证: Apache-2.0
- 标签: world-model, active-learning, confidence-estimation, robot-learning, robotwin
数据包内容
-
50-task prescreen package/
50 个任务的预筛选包(每任务一个代表性轨迹),包含 EVAC-v1 推理结果、置信度分数、风险图及 JSON 元数据。内含task_prescreen_package.tar和task_prescreen_summary.json。 -
EVAC-v2 training - inference/
C3 均值风险采集选中的全部轨迹(7298 条,覆盖权重变体并集)的预计算 EVAC-v1 推理输出。按任务分片(50 个 tar),含预测帧、潜变量、动作及元数据,并附none_selection_summary.json。 -
EVAC-v2 training - dense confidence/
同一批选中轨迹(7277 条,剩余 21 条预筛选代表在预筛选包中评分)的密集置信度/风险输出,用于置信度引导重训练(帧/帧+块加权)。含c3_mean_risk_selected_scores.tar和选择概览 JSON。 -
Baseline selection - v1 inference results/
其他采集基线(GVL、LRMS、PRM-judge、random、RoboMeter-pref、RoboMeter-prog、RoboReward,均使用均值风险聚合)选中的轨迹的 EVAC-v1 推理输出,排除已在 C3 包中的轨迹。每基线一个 tar 包和对应选择概览 JSON。 -
Baseline weighting - v2 frame-scoring data/
附加权重基线实验的帧级评分产物,包含baseline_weighting_v2_frame_scoring.tar(含 lrms/、prm_judge/、robometer_pref/、robometer_prog/ 目录及 JSONL/JSON 文件)。 -
YOLO RoboTwin2.0 annotations/
根据 RoboTwin2.0 动作条件估计的机械臂夹爪轨迹标签(YOLO 检测格式)。含yolo_annotations.tar(标签、data.yaml、清单、投影报告及可视化文件)。注意:源图像(94 GB)未重新分发,需将data.yaml指向本地 RoboTwin2.0 帧。 -
Evaluation tables and bootstrap JSON/
含pooled_mean_risk_episode_level.json,提供均值/逐种子指标及轨迹级均值风险比较的合并配对 bootstrap 统计。
规范与匿名化
- tar 归档未压缩(
tar -x解压),每 RoboTwin 任务一个分片。 - 所有 JSON/YAML/CSV 均经过路径清理:绝对路径替换为仓库相对路径或
<DATA_ROOT>/…占位符,无用户名、主机名或机器特定路径残留,数值数组和图像保持不变。 - 轨迹 ID 遵循 RoboTwin2.0 格式(
<task>_aloha-agilex_randomized_500_epNNN),可通过<DATA_ROOT>路径从公开 RoboTwin2.0 发布中恢复原始轨迹。
配套资源
- 配套模型检查点:
anonymous89793/ConfAL-WM模型仓库。
数据集地址:https://huggingface.co/datasets/anonymous89793/ConfAL-WM-Dataset





