DAVIAN-Robotics/RoboLab-BananaInBowl-30cm-relrate-k1c0-det-gripbin-260716
收藏资源简介:
该数据集使用LeRobot创建,包含RoboLab(Isaac Lab)BananaInBowl @30cm任务的轨迹数据,这些数据来自RFCL x RoboLab强化学习策略,并已转换为LeRobot v3.0格式,适用于pi0.5(pi05)监督微调(SFT)。具体细节包括:策略为relrate k1-c0(速率限制的相对关节动作;每轴上限=演示数据p99(|delta|) x 1.0),采样方式为确定性采样,动作空间记录为绝对关节目标(droid坐标系,j7偏移为-pi/4),夹爪状态为二进制{0,1}(1表示闭合)。成功定义为K-hold稳定条件(success_hold_steps=15),初始化随机化距离为30cm。相机配置包括over_shoulder_left_camera(外部)和wrist_cam,在pi05 SFT中分别映射到base_0_rgb和left_wrist_0_rgb。统计信息包含q01/q99分位数(用于pi05分位数归一化),视频为CFR归一化。
This dataset was created using LeRobot. RoboLab (Isaac Lab) BananaInBowl @30cm — rollouts from an RFCL x RoboLab RL policy, converted to LeRobot v3.0 for pi0.5 (pi05) SFT. Provenance: Policy: relrate k1-c0 (rate-limited relative joint action; per-axis cap = demo p99(|delta|) x 1.0). Sampling: det. Action space as recorded: ABSOLUTE joint targets (droid frame, j7 offset -pi/4). Gripper: binary {0,1} (1 = close). Success: K-hold settle, success_hold_steps=15, success-as-truncation. Init randomization: 30cm. Cameras: over_shoulder_left_camera (exterior) + wrist_cam. For pi05 SFT these map to base_0_rgb / left_wrist_0_rgb. Stats include q01/q99 (required by pi05 quantile normalization). Videos are CFR-normalized.




