File size: 5,211 Bytes
1b8ae96
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
---
license: mit
task_categories:
  - robotics
tags:
  - lerobot
  - so101
  - act
  - reinforcement-learning
  - exploration
  - outcome-labeled
  - manipulation
size_categories:
  - n<1K
---

# so101_pick_diverse_objects_rl_round1

## Dataset Description

This dataset contains **exploration rollout episodes with outcome labels** collected on the SO-101 robot arm pair using [`rl_rollout_act.py`](https://github.com/TakuyaHiraoka/LeRobot-Playground).

The collection procedure is as follows:

1. An ACT policy (trained on HIL-corrected data; see [so101_pick_diverse_objects_hil_round1](https://huggingface.co/datasets/TakuyaHiraoka/so101_pick_diverse_objects_hil_round1)) drives the follower arm autonomously.
2. Gaussian noise is injected into the ACT **latent space** (`z`) at each step to encourage exploration beyond the nominal policy distribution.
3. At the end of each episode, the human operator labels the outcome as **success** or **failure**.
4. Every saved episode includes `next.reward` and `next.done` columns: `0 / 0` for all frames, with the final frame of a success episode set to `reward = 1, done = 1`.

This format makes the dataset directly compatible with **REINFORCE-style policy updates** (filtering to success episodes only, via `train_act_filtered.py`) and **binary reward classifier training** (via `train_reward_classifier.py`).

## Task

**Pick diverse objects** — the robot arm picks up a variety of objects placed on a tabletop. Multiple task prompts are supported (e.g., *"Pick the red cube."*, *"Pick a pen."*); the task string is assigned at the start of each episode.

## Key Differences from the HIL Dataset

| Property | [hil_round1](https://huggingface.co/datasets/TakuyaHiraoka/so101_pick_diverse_objects_hil_round1) | **rl_round1** (this dataset) |
|---|---|---|
| Data source | Human intervention corrections | Autonomous rollouts with noise |
| Human role | Takes over during failures | Labels outcome at episode end |
| Exploration mechanism | Human guidance | Gaussian latent-space noise (`--noise_std`) |
| Reward signal | Not included | `next.reward`, `next.done` per frame |
| Primary use | Policy fine-tuning (BC) | REINFORCE update / reward classifier |

## Related Resources

| Resource | Link |
|---|---|
| **Seed HIL dataset (Round 1)** | [TakuyaHiraoka/so101_pick_diverse_objects_hil_round1](https://huggingface.co/datasets/TakuyaHiraoka/so101_pick_diverse_objects_hil_round1) |
| **Policy trained on HIL data** | [TakuyaHiraoka/act_so101_pick_diverse_objects](https://huggingface.co/TakuyaHiraoka/act_so101_pick_diverse_objects) |
| **Full pipeline & scripts** | [TakuyaHiraoka/LeRobot-Playground](https://github.com/TakuyaHiraoka/LeRobot-Playground) |

## Hardware

- **Robot:** SO-101 leader–follower arm pair
- **Camera:** USB or IP/MJPEG stream (640 px wide)
- **Framework:** [LeRobot](https://github.com/huggingface/lerobot) `== 0.4.4`

## Data Collection Procedure

The dataset was collected using `rl_rollout_act.py` from [LeRobot-Playground](https://github.com/TakuyaHiraoka/LeRobot-Playground).

Key keyboard controls during recording:

| Key | Action |
|---|---|
| `s` | Save current episode as **success** |
| `f` | Save current episode as **failure** |
| `r` | Discard and re-record |
| `q` | End the session |

On timeout, the terminal always prompts for an outcome label — every saved episode is guaranteed to have a `success` or `failure` annotation.

Example launch command:

```bash
python rl_rollout_act.py \
    --follower_port /dev/ttyACM1 --follower_id <follower-id> \
    --leader_port   /dev/ttyACM0 --leader_id   <leader-id> \
    --camera_url    "http://<camera-host>:8080" \
    --policy_path   <path-to-pretrained-act> \
    --repo_id       TakuyaHiraoka/so101_pick_diverse_objects_rl_round1 \
    --tasks         "Pick a pen." \
    --num_episodes 20 --episode_time_s 20 --reset_time_s 5 --fps 30 \
    --noise_std 0.05 --resume
```

## Downstream Usage

### REINFORCE-style ACT re-training (success filtering)

```bash
python train_act_filtered.py \
    --dataset_repo_id TakuyaHiraoka/so101_pick_diverse_objects_rl_round1 \
    --base_policy     <path-to-act-v1> \
    --output_dir      outputs/train/act_v2 \
    -- \
    --steps 30000 --batch_size 64 --policy.optimizer_lr 1e-5
```

`train_act_filtered.py` scans for episodes where `max(next.reward) > threshold` and passes only those to `lerobot-train`. For binary reward with no advantage, this is equivalent to a REINFORCE update:

```
∇θ J = E_τ [ ∇θ log π(τ) · R(τ) ]
```

### Binary reward classifier training

```bash
python train_reward_classifier.py \
    --dataset_repo_id TakuyaHiraoka/so101_pick_diverse_objects_rl_round1 \
    --output_dir      outputs/reward_clf/round1 \
    --epochs 20 --batch_size 64
```

Trains a ResNet-18 head to predict `P(success | s)` per frame. Validation is held out at the episode level to prevent data leakage.

## Intended Use

This dataset is intended for:

- **REINFORCE / filtered BC** — re-train the ACT policy on success episodes only.
- **Learned value / reward functions** — train a binary success classifier as an approximate value function.

## License

[MIT](https://opensource.org/licenses/MIT)