novastar112/pusht_96_int1_visual_nomarker_allstep_thinking_trickiness_cot

该数据集源自novastar112/pusht_96_int1_visual_nomarker中的成功PushT视觉无标记轨迹，包含从初始移动到最终停止动作的完整成功轨迹。主要文件包括训练集（500,000行）和测试集（200行）。消息格式中，用户回合包含PushT提示文本和当前图像项（损失标记为false），图像项存储前一张图像、当前图像和下一张图像（已验证前一张图像与当前图像相同）。每个非停止移动步骤中，选择覆盖变化最大的一步，并生成四个想象的一步分支图像（损失标记为true）。其他非停止助手回合使用固定的简单思考模板，最终助手回合使用固定的停止思考模板并输出(stop, stop)。关键步骤策略包括回放源轨迹、选择覆盖变化最大的非停止移动步骤、使用48像素的PushT世界扰动半径，并优先围绕计划目标进行运动帧扰动（如计划目标、较短目标、较远目标、左路径、右路径），如果运动帧试验集不具对比性，则回退到目标帧扰动。最终保留计划目标加上三个最强对比性扰动，共生成四个分支图像。This dataset is derived from successful PushT visual-nomarker trajectories in novastar112/pusht_96_int1_visual_nomarker. Each row contains one full successful trajectory from the first move through the final stop action. Main files include a training set (500,000 rows) and a testing set (200 rows). In the message format, each user turn consists of the PushT prompt text plus one current image item with loss=false, storing image_prev, image, and image_next (with image_prev == image validated for every step). One selected max-coverage-jump non-stop move step contains four imagined one-step branch images with loss=true. All other non-stop assistant turns use a fixed trivial thinking template, and the final assistant turn uses a fixed stop thinking template and outputs (stop, stop). The key-step policy involves replaying the source trajectory and selecting the non-stop move step with the largest one-step coverage change, using a 48 px PushT-world perturbation radius, preferring motion-frame perturbations around the planned target (planned, shorter_target, farther_target, left_path, right_path), and falling back to goal-frame perturbations only if the motion-frame trial set is not contrastive. It keeps the planned target plus the three strongest contrastive perturbations, for four branch images total.