Hoshipu/ARCHIEVED-behavior-1k-embodiedAI-rollouts-hq-use

BEHAVIOR-1K Task-0 — Demos + X-VLA Rollouts是一个合并的数据集，专门针对BEHAVIOR-1K任务0（“打开客厅桌子上收音机接收器”）。它结合了200个人类远程操作演示（来自HuggingFace上游数据集behavior-1k/2025-challenge-demos）和259个X-VLA策略rollouts（通过OmniGibson在相同场景中本地收集，使用Hoshipu/xvla-behavior1k-v20的task0-60k检查点）。所有rollouts均为成功完成的任务。数据集遵循LeRobot v2.1布局，采用基于任务的分块结构（data/task-{NNNN}/episode_{IIIIIIII}.parquet），便于现有BEHAVIOR-1K/LeRobot消费者直接使用。数据包括机器人的状态观察（256维）、动作（23维）、时间戳、任务信息等，以及RGB视频流。用途包括DAgger式模仿学习、行为克隆消融、逆运动学先验训练和新策略基准测试。数据集总共有459个episodes（200个演示 + 259个rollouts），涵盖R1Pro双足人形机器人的动作和状态数据。

BEHAVIOR-1K Task-0 — Demos + X-VLA Rollouts is a merged dataset for BEHAVIOR-1K Task 0 (turning on the radio receiver on the table in the living room). It combines 200 human-teleoperated demonstrations from the upstream HuggingFace dataset behavior-1k/2025-challenge-demos and 259 X-VLA policy rollouts collected locally on the same scenes via OmniGibson using the Hoshipu/xvla-behavior1k-v20 task0-60k checkpoint, all of which are successful completions. The dataset follows the LeRobot v2.1 layout with task-based chunking (data/task-{NNNN}/episode_{IIIIIIII}.parquet), enabling seamless integration with existing BEHAVIOR-1K/LeRobot consumers. It includes robot state observations (256-D), actions (23-D), timestamps, task info, and RGB video streams. The dataset is designed for applications such as DAgger-style imitation learning, behavior cloning ablations, inverse-kinematics priors training, and benchmarking new policies. In total, there are 459 episodes (200 demos + 259 rollouts), covering action and state data for the R1Pro dual-arm humanoid robot.