# Jason's multi_task_dit — SC1 Final Checkpoint

**成功率: 9/9 (100%) on SO-101 real robot (2026-04-17)**
Policy: `multi_task_dit` (lerobot 0.5.1 built-in, ~277M params)

## 環境

```bash
conda activate a   # Jason's env on idlab2
# lerobot 來自 /home/kunhsiang/jdk/lib/lerobot/
# CLIP weights: /home/kunhsiang/jdk/hf_weights/clip-vit-base-patch16/
```

## 訓練指令（完整復現）

```bash
cd /home/kunhsiang/jdk
CUDA_VISIBLE_DEVICES=1 lerobot-train \
  --dataset.root=datasets/sc1_newcolor \
  --dataset.repo_id=YOUR_DATASET \
  --output_dir=outputs/train/multi_task_dit_repro \
  --policy.type=multi_task_dit \
  --policy.device=cuda \
  --policy.horizon=32 \
  --policy.n_action_steps=24 \
  --policy.objective=diffusion \
  --policy.noise_scheduler_type=DDIM \
  --policy.num_train_timesteps=100 \
  --policy.num_inference_steps=20 \
  --policy.clip_sample=true \
  --policy.clip_sample_range=1.0 \
  --policy.hidden_dim=512 \
  --policy.num_layers=6 \
  --policy.num_heads=8 \
  --policy.repo_id=jedeka30/grasp_box \
  --wandb.enable=true \
  --wandb.project=act \
  --steps=50000 \
  --save_freq=5000 \
  --batch_size=24 \
  --num_workers=8 \
  --eval_freq=2000
```

## 超參數摘要

| 參數 | 值 |
|------|-----|
| Policy | multi_task_dit |
| Objective | diffusion (DDIM) |
| horizon | 32 |
| n_action_steps | 24 |
| **num_inference_steps** | **20** (關鍵，比 10 好很多) |
| num_train_timesteps | 100 |
| hidden_dim | 512 |
| num_layers | 6 |
| num_heads | 8 |
| batch_size | 24 |
| steps | 50,000 |
| Dataset | sc1_newcolor (3色: red/blue/green, ~80 episodes) |

## Eval 指令

```bash
cd /home/kunhsiang/jdk

# SC1: red/green/blue 三色（自動循環）
bash 3kh.sh ckpts/final_ckpts/pretrained_model

# SC2/3: on/outside the plate
bash 2test.sh ckpts/final_ckpts/pretrained_model
```

## 硬體設定

| 設備 | 值 |
|------|-----|
| Robot | SO-101 follower |
| Port | /dev/ttyACM1 |
| GPU | CUDA_VISIBLE_DEVICES=1 |
| cam_front | /dev/video2 (640×480, 30fps, MJPG) |
| cam_gripper | /dev/video0 (640×480, 30fps, MJPG) |
| cam_top | /dev/video4 (640×480, 30fps, MJPG) |
| episode_time_s | 90 |

## 評測結果

| Task | SR | Date |
|------|-----|------|
| grasp sc1 (9 trials: red×3 / green×3 / blue×3) | **9/9 (100%)** | 2026-04-17 |
| grasp sc4 (put on plate) | 2/6 (33%) | 2026-05-04 |

## 注意

- lerobot 的 `multi_task_dit` 有 offline CLIP hack: 讀 `/home/kunhsiang/jdk/hf_weights/clip-vit-base-patch16/`
  標準 lerobot 0.5.1 會從 HF 下載 `openai/clip-vit-base-patch16`（需網路）
- `TRANSFORMERS_OFFLINE=1` + `HF_DATASETS_OFFLINE=1` 在訓練時設定（離線模式）
- 訓練完後 `make reset` 會把 robot 歸位（Makefile 在 jdk 根目錄）

## 相關 Repo

- Dataset: `kunhsiang/jason-sc1-newcolor-dataset`
- Eval scripts: 此 repo 內 `scripts/` 目錄