sprout-real-3task / README.md
Jeongeun's picture
card: human_ghost_r29
39393e7 verified
|
Raw History Blame Contribute Delete
5.47 kB
metadata
license: other
task_categories:
  - robotics
tags:
  - robot-learning
  - imitation-learning
  - real-robot
  - sprout

sprout-real-3task

Teleoperated demonstrations on the Sprout humanoid (real robot), 3 tasks x 20 episodes, in the training-ready 29-D format used by cosmos_policy (cosmos_predict2p5_2b_sprout_3task_pv0_*, guide: docs/real-robot-context-free.md).

task folder episodes instruction arms
sprout_cup_to_drawer 20 Pick up the cup and put it in the drawer right; left written at the reset pose
sprout_open_drawer 20 Open the drawer left; right written at the reset pose
sprout_cup_close_drawer 20 Pick up the cup and close drawer both

60 episodes, 10,358 frames at 15 Hz (11.5 min), one training window per frame (10,358).

Layout

<task>/episode_<N>.h5
    images/front_1      (T,) vlen uint8 JPEG 224x224   ego view (ZED left)   <- the camera the configs use
    images/center_top   (T,) vlen uint8 JPEG 224x224   ZED right (not used)
    state29             (T, 29) float32                FK of the measured joints
    abs_action29        (T, 29) float32                FK of /motor_control/joints/applied_command   <- target
    attrs: n_frames, fps=15, instruction, src (recording), start_s/end_s, idle_arm
manifest/
    windows.npz  pairs.npz  tasks.json  t5_embeddings.pkl (T5-11B, 3 instructions)
    state29_stats_pctl.json  abs_action29_stats_pctl.json   (1/99 percentile min-max)

29-D vector, all poses in torso_link (x front, y left, z up):

0:9    head9   torso velocity command [vx, vy, vz=0, rot6d(Rz(vyaw / 15))]; standing still = identity
9:18   Leef9   left  <side>_gripper_lower_link: xyz + rot6d (first two columns of R)
18     Lgrip   0 open .. 1 closed
19:28  Reef9   right, as left
28     Rgrip

In the cup and drawer tasks the arm the task does not use is written at the reset pose in both state and action (idle_arm attr): stand pose with the shoulder roll 0.10 rad further out, wrists 0, gripper closed (1.0). The images are not changed. Episodes are the last take of each task number, trimmed by hand to the task.

Human ghost demos (in-context post-train)

human_ghost/ holds one human demonstration per robot episode of the two single-arm tasks (40 files, 349 MB), used by cosmos_predict2p5_2b_sprout_icl_c32* (guide: docs/real-robot-icl.md in the code repo). Each is a "ghost": a person wearing a Quest 3 followed that robot episode, replayed as a ghost hand in VR, at the robot's speed, and the recording is trimmed to the robot episode's length, so human frame t is robot frame t.

human_ghost/<task>/episode_<N>.h5            the ghost of <task>/episode_<N>.h5  (sprout_cup_to_drawer, sprout_open_drawer)
    images/front_1   (T,) vlen uint8 JPEG 640x480   Quest 3 left passthrough camera (the loader resizes to 224x224)
    state57          (T, 57) float32                EgoVerse state57: waist frame at frame 0, wrist eef + 15 hand keypoints
    attrs: n_frames (= the robot episode's), fps=15, human_episode, robot_task, robot_episode
manifest/
    human_state57_stats.json   per-dim mean -/+ 4 std of state57 over the 40 ghosts (key "quest")
    human_ghost.json           {human_episode: "<task>/episode_<N>"}
    human_pairs.npz            retrieval pairs, robot 32-frame windows (every 5 frames) -> top-8 human 32-frame windows
                               of other ghosts (wrist-path DTW in the body frame + gripper timing); used only by the
                               _mix30 arm

The tag-0 task variants were recorded; the pick-and-place + close task (sprout_cup_close_drawer) has no ghosts.

Retargeted ghosts (human_ghost_r29/)

The same 40 ghosts converted to the robot's own representation, used by cosmos_predict2p5_2b_sprout_icl_c32_r29* (_r29abs_cfaug[_within|_base], _r29rel_cfaug[_within|_base]; guide: docs/real-robot-icl.md). Made by quest_capture/pc/retarget.py in the code repo; human_ghost/ is unchanged.

human_ghost_r29/<task>/episode_<N>.h5        a copy of human_ghost/<task>/episode_<N>.h5 plus
    state29          (T, 29) float32   robot state29 layout: head9 = base velocity command [vx, vy, 0, rot6d(Rz(vyaw/15))],
                                       working-arm eef9 (torso frame) + gripper, the other arm at its reset pose
    state29abs       (T, 29) float32   the same with head9 = the torso pose in the start frame [x, y, 0, rot6d(Rz(yaw))]
    attrs: retarget_side, grip_from (the robot episode whose gripper is used)
manifest/
    human_anchor.json          per human episode: the ghost guide's anchor (= the robot's start frame) in its waist frame
    retarget_calib.json        the fitted constants (per-side hand offset and rotation, base rules)
    ghost_match_r29.json       {"<task>/<episode>": [h(t) for every robot frame t]}: the demo frame the inference
                               retrieval picks (retargeted EEF + gripper + base + position, band +8..+96 per chunk,
                               held while the robot is still); the r29 arms place the training demo window there

Retargeting: the hand path is placed in the robot's start frame by the guide anchor; the base holds still and moves only when the person does (head yaw / position), as one robot-speed turn or step (60 deg/s, 0.38 m/s) at the person's turn / step; the arm is the hand in that torso frame; the gripper is the followed robot episode's.