File size: 3,090 Bytes
cd15d41 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 | ---
license: apache-2.0
tags: [robotics, lerobot, real-robot]
---
# shell_game — delta EEF, per-frame subtask
Cups hide a red block, the cups are shuffled by hand, and the arm lifts the cup the block ended
under. 45 episodes, 18,442 frames.
Delta EEF, 10-dim. Six sentences, assigned per frame:
```
observe 6,051 frames
uncover the left cup 2,203
uncover the middle cup 1,517
uncover the right cup 1,848
move the red cube into the basket 6,823
```
Labels are derived from state, never from the video: the gripper separates cup (closure
0.120-0.200) from cube (0.440-0.460) with no overlap and exactly two grasps in 45 of 45
episodes, and the cup is read off cartesian y, which is 3-modal. They agree with the human
collection sheet on 44 of 45; the one disagreement is an episode where the operator lifted the
wrong cup.
Which cup to uncover cannot be derived from the current frame — the shuffle has already
happened and every cup looks alike. That is the point of the task, and it is why a planner
trained on this needs memory rather than perception.
Converted from the recorder's DROID-schema `teleoperation.h5`, one episode per directory, with
the videos re-encoded all-intra. The raw h5 is the source of record and lives at
`ghkim-rlwrld/shell_game`; anything that needs a different action representation or a fresh label
pass has to go back to it, not to this tree.
Cameras are three stereo pairs and only the left eye is shipped here, which is what a policy
reading three streams actually loads:
```
11845904 wrist looks down over the gripper
37149196 cam_ext_left external, to the robot's left
34022131 cam_ext_right external, to the robot's right
```
Control rate is 9.2 Hz measured. The mp4 containers say 10 fps and are wrong.
### Naming
`<task>_<action space>_<language>`. An earlier set of names took the action space from one
dataset and the language from another, so they could not be compared side by side.
| | action | language |
|---|---|---|
| `shell_game_joint_global` | 7 joints + gripper (8) | one sentence per episode |
| `shell_game_eef_global` | delta EEF, xyz + rot6d + gripper (10) | one sentence per episode |
| `shell_game_eef_subtask` | delta EEF (10) | per-frame subtask |
| `solve_equation_eef_subtask` | delta EEF (10) | per-frame subtask |
### Why rot6d and not rpy
This recorder's roll sits at +-pi and flips sign 269 times across the 45 shell_game episodes.
Regressing the raw angle puts a 2*pi step in the target at every flip. Unwrapping and moving to
the first two columns of the rotation matrix leaves 0 discontinuities and takes the roll channel
from std 3.08 to 0.039. The delta mask is `(9, -1)`: position and rotation relative to the
current state, gripper absolute.
A runtime that wants 7-dim xyz+rpy can recover it exactly — Gram-Schmidt the two 3-vectors, take
the third by cross product, then `as_euler("xyz")`, verified to 4.44e-16 over 41 episodes. Doing
that at inference is safe; training on rpy is not, for the reason above.
|