| --- |
| tags: |
| - LeRobot |
| - droid |
| task_categories: |
| - robotics |
| --- |
| |
| # DROID 1.0.1 with delta-EE actions and gripper-event timings |
|
|
| `lerobot/droid_1.0.1` with added per-frame features: |
|
|
| - `action.delta_ee` — `[x, y, z, rx, ry, rz, gripper]` (float32): per-step |
| delta EE command: position delta of `action.cartesian_position` vs |
| `observation.state.cartesian_position`, and the relative rotation |
| `R_cmd * R_state^-1` (world frame) as an axis-angle vector (gimbal-safe, unlike |
| elementwise rpy subtraction); gripper is the absolute command (0..1, 1 = close). |
| - `observation.extrinsics.static1` / `static2` / `wrist1` — `[pos(3), quat xyzw(4)]` |
| (float32) camera poses in the robot base frame (static1 = exterior_1_left, |
| static2 = exterior_2_left, wrist1 = wrist_left), converted from |
| `camera_extrinsics.*` (xyz + extrinsic-xyz rpy). The wrist pose is live per frame |
| (EE pose composed with the episode's rigid mount offset); the source column |
| freezes it at episode start. |
| - `observation.gripper.time_to_close` / `time_to_open` — hindsight seconds until the |
| gripper command next flips (threshold 0.5, runs < 3 steps debounced). |
| - `observation.gripper.time_since_close` / `time_since_open` — seconds since the |
| previous flip. Frames with no such flip in the episode hold the sentinel `10000`. |
| |
| The flat `action` column holds the delta-EE values and the flat `observation.state` |
| column holds the EE state `[x, y, z, roll, pitch, yaw, gripper]`; the joint-space |
| originals remain in `action.joint_position` / `observation.state.joint_position`. |
| |
| Generated by `lerobot_policy_framepick/scripts/augment_droid_delta_ee_gripper_events.py`. |
| |
