File size: 2,280 Bytes
effc88c
1c23c72
 
effc88c
1c23c72
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
feec153
 
 
 
1c23c72
feec153
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
---
license: mit
tags: [robotics, franka, maniskill, simulation, lerobot]
---
# manip_dataset: FR3 simulated manipulation demos (8 tasks)

Scripted successful demonstrations of a Franka Research 3 arm in ManiSkill3, recorded for fine-tuning pi0.5.
Only successful episodes are kept. 3200 episodes in total.

| task (env id) | episodes | prompt |
|---|---|---|
| BlockBalance-v1 | 400 | place the beam on the red cube, then put one cube on each side of the beam so that it stays balanced |
| CarryStackToShelf-v1 | 400 | move the box onto the shelf while keeping the small cube on top |
| LoadAndMovePlate-v1 | 400 | place the blue cube on the plate, then move the plate with the cube to the red zone |
| PourMarbles-v1 | 400 | pick up the cup with the marbles, pour them into the empty cup and put the cup back on the red square |
| PutBlocksInBin-v1 | 400 | put all the cubes in the bin |
| SlidePuckToTarget-v1 | 400 | slide the red puck so it stops on the green target |
| SortThenStackFR3-v1 | 400 | drag the blue cube into the red zone, then stack the yellow cube on it |
| StackCubeFR3-v1 | 400 | stack the red cube on top of the green cube |

- Cameras: `front` (external right), `wrist`, `external_left` (opposite external view); 224x224 RGB.
- Rates: control 20 Hz; images and the export at 10 Hz.
- Randomization: object positions / yaws / sizes per task (all objects at least 15 cm from the table edges), robot start-pose
  joint noise (sigma 0.1 rad, clipped at 2 sigma). Domain randomization (lighting, camera pose, image noise, latencies) is off.
- Controller: `pd_ee_target_pose` (Cartesian end-effector target pose + gripper width).

## Layout

- `raw/<EnvId>/`: all 400 episodes of a task (seeds 0..447 from the first recording, 1000.. from the second; same settings).
  - `scripted_pd_ee_target_pose.{h5,json}` (seeds 0..) and `scripted_pd_ee_target_pose_seed1000.{h5,json}` (seeds 1000..):
    full 20 Hz trajectories (qpos, qvel, joint torques, TCP pose, actions, success) with 10 Hz images of the 3 cameras,
    ManiSkill trajectory format.
  - `export_oopsie_10hz/sim_<seed>.h5` + `sim_<seed>_<camera>.mp4`: 10 Hz Oopsie export (Cartesian states / targets, videos).
  - `recorder_seed0.log`, `recorder_seed1000.log`: every attempt with its seed and outcome.