Datasets:

Candlest
/

Robot-Use-Agent-SFTData

license: apache-2.0
language:
  - en
  - zh
pretty_name: Robot-Use-Agent-SFTData
size_categories:
  - 10K<n<100K
task_categories:
  - image-text-to-text
  - robotics
tags:
  - robotics
  - embodied-ai
  - agent
  - sft
  - code-as-policy
  - multi-turn
  - llama-factory
  - axolotl
  - qwen2-vl

Robot-Use-Agent-SFTData

A comprehensive multi-turn multimodal dataset for supervised fine-tuning (SFT) of vision-language embodied robot agents, converted from the standardized robot-control-trace/v1 format.

Compatible out-of-the-box with Hugging Face Datasets, LLaMA-Factory, Axolotl, and CaPtain-SFT.

Dataset Statistics

Metric	Value
Total Episodes	759 episodes
Total Interaction Steps	34,343 steps
Camera Frames	113,609 images (5.6 GB)
Cameras	Multi-view (`cam_head`, `cam_left_wrist`, `cam_right_wrist`)
Format	Multi-turn conversational `messages` with tool calls & CoT reasoning
Benchmarks & Sources	RoboDojo, RoboTwin (92 manipulation tasks), Local Codex ARX X5 real rollouts, Remote audit clusters

Files & Structure

Robot-Use-Agent-SFTData/
  ├── sft.jsonl              # Core SFT training rows (ShareGPT / CaPtain multimodal messages)
  ├── dataset_info.json      # LLaMA-Factory direct registration config
  ├── manifest.json          # Master metadata catalog of all 759 episodes
  ├── images_manifest.tsv    # TSV index of all 113,609 image files
  ├── images.tar             # Complete image archive (5.6 GB)
  ├── unpack_images.py       # One-click image unpacker
  └── README.md              # Dataset card & documentation

Quick Start & Usage

1. Download & Unpack Images

# Clone repository
git clone https://huggingface.co/datasets/Candlest/Robot-Use-Agent-SFTData
cd Robot-Use-Agent-SFTData

# Unpack camera observations (takes ~15 seconds)
python3 unpack_images.py

2. Load with Hugging Face Datasets

from datasets import load_dataset

dataset = load_dataset("json", data_files="sft.jsonl")
print("Total training episodes:", len(dataset["train"]))

first_sample = dataset["train"][0]
print("Episode ID:", first_sample["id"])
print("Total Turns:", len(first_sample["messages"]))
print("Chain of Thought:", first_sample["messages"][2]["reasoning_content"])

3. Fine-tuning with LLaMA-Factory (Qwen2.5-VL / Qwen3.5-VL)

{
  "captain_robot_sft": {
    "file_name": "sft.jsonl",
    "formatting": "sharegpt",
    "columns": {
      "messages": "messages",
      "images": "images",
      "tools": "tools"
    }
  }
}

Launch training:

llamafactory-cli train \
    --stage sft \
    --do_train \
    --model_name_or_path Qwen/Qwen2.5-VL-7B-Instruct \
    --dataset captain_robot_sft \
    --dataset_dir /path/to/Robot-Use-Agent-SFTData \
    --template qwen2_vl \
    --finetuning_type lora \
    --output_dir ./output/robot_vlm_lora \
    --cutoff_len 4096 \
    --learning_rate 1e-4 \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 8

Row Schema

Each line in sft.jsonl follows standard multi-turn tool-use format:

{
  "id": "robodojo_fold_clothes_L1#full",
  "images": [
    "images/robodojo_fold_clothes_L1_0000_cam_head.jpg",
    "images/robodojo_fold_clothes_L1_0000_cam_left_wrist.jpg",
    "images/robodojo_fold_clothes_L1_0000_cam_right_wrist.jpg"
  ],
  "tools": [{
    "type": "function",
    "function": {
      "name": "robot_controller",
      "description": "Execute robot primitive actions, motion commands, or skill policies on the robot.",
      "parameters": { ... }
    }
  }],
  "messages": [
    {"role": "system", "content": "You are an embodied robotic agent..."},
    {"role": "user", "content": [
      {"type": "text", "text": "Task: Fold the clothes neatly.\nInitial camera observations:"},
      {"type": "image", "image": "images/robodojo_fold_clothes_L1_0000_cam_head.jpg"},
      {"type": "image", "image": "images/robodojo_fold_clothes_L1_0000_cam_left_wrist.jpg"},
      {"type": "image", "image": "images/robodojo_fold_clothes_L1_0000_cam_right_wrist.jpg"}
    ]},
    {"role": "assistant", "content": "Executing: fold_cloth_left", "reasoning_content": "Plan: Approach the corner of the cloth...", "tool_calls": [...]},
    {"role": "tool", "tool_call_id": "call_robodojo_fold_clothes_L1_step_0000", "content": [{"type": "text", "text": "ok"}]}
  ],
  "meta": {
    "episode": "robodojo_fold_clothes_L1",
    "benchmark": "robodojo",
    "task": "fold_clothes",
    "total_steps": 15,
    "n_images": 45,
    "supervise_turns": [2, 5, 8, ...],
    "host_final_success": true
  }
}

Citation & License

License: Apache-2.0
Underlying Traces Schema: robot-control-trace/v1