File size: 2,737 Bytes
887ac18
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
tags:
- lerobot
- dataforge
- robotics
- bi_so101_follower
task_categories:
- robotics
---

# robocollect/so-101-6

A [LeRobot](https://github.com/huggingface/lerobot) dataset for the
**so101_new_calib**, built by DataForge from human
demonstration video. 15 episodes, 10152 frames at 30 fps.

## Columns

| Column | What it is |
|---|---|
| `observation.images.top` | The operator's video, re-encoded |
| `observation.state` | Joint positions in the robot's dataset units (its URDF's angles), the gripper in its own values |
| `action` | The commanded joint positions, same units |

12 values, `left_arm_shoulder_pan.pos` … `right_arm_gripper.pos`, robot type `bi_so101_follower`: the left arm is the hand that worked on the left of the picture, and both arms share one clock.
Each robot is read from its own URDF: no calibration file. The gripper runs
between the robot's own dataset values for open and closed.

## How it was made

Nobody teleoperated a robot to record this. Each episode is a phone video of a
person doing the task with their own hand, retargeted onto the so101_new_calib by
`hand2robot-1.0`: a 3D hand reconstruction, then
differential IK onto its URDF, then retiming inside its joints' speed
limits.

Episodes are on the **robot's** clock, which runs about 1.6 to 2 times slower
than the human's, so video frames repeat where the arm moves slowly. The
travel to and from the calibration pose is not included.

Only clips that passed both an automatic content check and a trajectory
quality gate are here. The gate rejects, among other things, clips where the
manipulated object was tracked in too few frames, where too much of the hand
path lies outside the arm's reach.

## What to know before you train on it

- **`observation.state` is not measured.** It is the previous commanded
  position: the IK trajectory, assuming the servos track it perfectly. No
  physical robot executed these trajectories.
- **The gripper closes when the object is held**, timed from the steps the
  automatic check named ("pick up ...", "place ..."). A hand around an object
  keeps thumb and index apart, so it is not read off the fingers. Episodes
  whose steps do not say follow the thumb-index distance.
- **The geometry is uncalibrated.** The camera focal length is approximated
  from the frame size and the hand scale is a population average, so expect a
  few centimetres of absolute error.
- **0 of 15 episodes** were solved against a real
  `lerobot-calibrate` file. Episodes without one assume symmetric joint
  ranges: fine to train on, wrong to replay on a particular arm.
- **Open loop.** The object is where the video said it was. A real arm needs
  the same layout relative to its base, or its own perception.