real01b_square_d2_final_trio_back50_ours_iql_s1final_heval_s2026070101_policy_rollouts

Real01b Square D2策略执行数据集（TsFile格式）是一个Apache TsFile转换版本，源自LeRobot v3.0 Franka策略执行数据集。该数据集包含时间序列数据，专注于机器人任务square_d2，采样率为15 fps，训练集包含50个episodes和11,977帧。数据集中包含116列，涵盖时间、episode索引、任务索引、帧索引、样本索引、步骤剩余、来源、干预、成功标志、有效性、奖励、完成标志、策略ID、回合ID、清单索引、任务几何参数（如螺母和钉子的位置和偏航角）、动作和状态测量（如夹爪位置和速度）、主要状态和动作向量、笛卡尔和关节动作向量、笛卡尔和关节观测向量，以及Franka扭矩和末端执行器力/力矩遥测数据。数据集排除了四个AV1相机流，但保留了与源视频对齐的帧索引和时间戳。转换过程包括列名规范化和向量扁平化处理，适用于机器人学习和策略评估研究。The Real01b Square D2 policy execution dataset (in TsFile format) is an Apache TsFile converted version, derived from the LeRobot v3.0 Franka policy execution dataset. It contains time-series data focused on the robot task square_d2, with a sampling rate of 15 fps, and the training set includes 50 episodes and 11,977 frames. The dataset comprises 116 columns, covering time, episode index, task index, frame index, sample index, steps remaining, source, intervention, success flag, validity, reward, completion flag, policy ID, round ID, manifest index, task geometry parameters (such as positions and yaw angles of nuts and pegs), action and state measurements (e.g., gripper position and velocity), main state and action vectors, Cartesian and joint action vectors, Cartesian and joint observation vectors, as well as Franka torque and end-effector force/torque telemetry data. The dataset excludes four AV1 camera streams but retains frame indices and timestamps aligned with the source video. The conversion process involves column name normalization and vector flattening, making it suitable for robot learning and policy evaluation research.