Datasets:
Update README.md
Browse files
README.md
CHANGED
|
@@ -11,6 +11,39 @@ configs:
|
|
| 11 |
- config_name: default
|
| 12 |
data_files: data/*/*.parquet
|
| 13 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
|
| 15 |
This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
|
| 16 |
|
|
@@ -29,15 +62,15 @@ This dataset was created using [LeRobot](https://github.com/huggingface/lerobot)
|
|
| 29 |
{
|
| 30 |
"codebase_version": "v2.1",
|
| 31 |
"robot_type": "2d pointer",
|
| 32 |
-
"total_episodes":
|
| 33 |
-
"total_frames":
|
| 34 |
"total_tasks": 1,
|
| 35 |
"total_videos": 0,
|
| 36 |
"total_chunks": 1,
|
| 37 |
"chunks_size": 1000,
|
| 38 |
"fps": 10,
|
| 39 |
"splits": {
|
| 40 |
-
"train": "0:
|
| 41 |
},
|
| 42 |
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
|
| 43 |
"video_path": "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4",
|
|
|
|
| 11 |
- config_name: default
|
| 12 |
data_files: data/*/*.parquet
|
| 13 |
---
|
| 14 |
+
# PushT Dataset
|
| 15 |
+
|
| 16 |
+
This dataset contains demonstrations for the PushT environment, a robotic manipulation task where an agent needs to push a T-shaped object onto a matching target surface.
|
| 17 |
+
Each episode is initialized with a **randomized** T-block and target position and orientation to ensure a more realistic scenario.
|
| 18 |
+
|
| 19 |
+
## Environment Details
|
| 20 |
+
|
| 21 |
+
The dataset was collected using the [gym-pusht](https://github.com/huggingface/gym-pusht) environment, which provides a simple 2D robotic pushing task.
|
| 22 |
+
|
| 23 |
+
- **Task**: Push the T-shaped gray block onto the T-shaped green target surface using the blue pointer.
|
| 24 |
+
- **Observation Space**: RGB images (96×96×3) and agent position (2D coordinates)
|
| 25 |
+
- **Action Space**: 2D coordinates for the agent position
|
| 26 |
+
- **Success Condition**: The T-shaped block overlaps sufficiently with the target surface
|
| 27 |
+
- **Randomized Goals**: This dataset uses a variation of the environment with `randomize_goal=True`, which randomizes the goal position for each episode, creating a more diverse and challenging dataset.
|
| 28 |
+
|
| 29 |
+
## Dataset Creation
|
| 30 |
+
|
| 31 |
+
This dataset was created through human demonstrations where a single human operator provided demonstrations of successful task completion.
|
| 32 |
+
|
| 33 |
+
### Dataset Composition
|
| 34 |
+
|
| 35 |
+
- **States**: RGB images and agent position coordinates
|
| 36 |
+
- **Actions**: 2D position commands for the agent
|
| 37 |
+
- **Rewards**: Sparse rewards based on task completion
|
| 38 |
+
- **Metadata**: Success flags and episode information
|
| 39 |
+
|
| 40 |
+
## Usage
|
| 41 |
+
|
| 42 |
+
This dataset is intended for training imitation learning and reinforcement learning policies for object manipulation tasks.
|
| 43 |
+
|
| 44 |
+
### Loading the Dataset
|
| 45 |
+
|
| 46 |
+
|
| 47 |
|
| 48 |
This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
|
| 49 |
|
|
|
|
| 62 |
{
|
| 63 |
"codebase_version": "v2.1",
|
| 64 |
"robot_type": "2d pointer",
|
| 65 |
+
"total_episodes": 101,
|
| 66 |
+
"total_frames": 9438,
|
| 67 |
"total_tasks": 1,
|
| 68 |
"total_videos": 0,
|
| 69 |
"total_chunks": 1,
|
| 70 |
"chunks_size": 1000,
|
| 71 |
"fps": 10,
|
| 72 |
"splits": {
|
| 73 |
+
"train": "0:101"
|
| 74 |
},
|
| 75 |
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
|
| 76 |
"video_path": "videos/chunk-{episode_chunk:03d}/{video_key}/episode_{episode_index:06d}.mp4",
|