uooon/reference_demo

🎬 GPT-GEAR Reference Demonstrations

🌐 English | 🌐 简体中文

Explore robot learning from demonstrations with GPT-GEAR
_{131 demonstrations · 4 simulation environments · Video, text, and action trajectories}

🤖 GPT-GEAR  ·  🧭 Overview  ·  📦 Contents  ·  🚀 Quick Start  ·  🗺️ Task Mapping  ·  💬 Community

---

🧭 Welcome

Welcome to the GPT-GEAR reference demonstration collection! This package brings together multi-view videos, textual steps, and action trajectories for RoboDojo, RoboTwin, LIBERO-GeoMotion, and LIBERO-Plus, making it easier to explore how vision-language models use examples to guide robot actions.

In GPT-GEAR, demonstrations become part of the model's context: the model can inspect an example, relate it to the current scene, and choose its next action using live observations and feedback. This is embodied in-context learning (ICL). The trajectories are references for the model to read, rather than commands automatically replayed on the robot.

Start by watching a demonstration or reading its steps, then try different input combinations in your own evaluations. We welcome reproduction notes, questions, and ideas from the community!

📦 What's included

Environment	Demonstrations	How they are used in GPT-GEAR
RoboDojo	35	34 same-name task references in the evaluation queue, plus spell_RoboDojo
RoboTwin	50	One per task, shared across clean / random configurations and evaluation seeds
LIBERO-GeoMotion	6	Shared within task families, covering 24 evaluation scenes
LIBERO-Plus	40	One per base task: 10 each from Spatial, Object, Goal, and LIBERO-10
Total	131	Video, text, and trajectory files for every demonstration

Each task directory follows the same layout:

reference_demo/
├── README.md
├── README.zh-CN.md
├── RoboDojo/<task>/
├── RoboTwin/<task>/
├── LIBERO-GeoMotion/task_<id>/
└── LIBERO-Plus/base_<id>/
    ├── steps.txt
    ├── trajectory.json
    ├── trajectory.jsonl
    └── videos/
        ├── videos.json
        └── *.mp4

File	What it contains	GPT-GEAR reference field
steps.txt	Textual steps and demonstration notes	reference.text
trajectory.json	Trajectory format, metadata, and file references	reference.trajectory
trajectory.jsonl	Detailed trajectory records, read through the descriptor	Referenced by trajectory.json
videos/	Camera-view MP4 files and the videos.json camera mapping	reference.video

Keep each task directory intact so the descriptor can resolve its trajectory and videos. Directory names and task IDs are case-sensitive.

RoboDojo, RoboTwin, and LIBERO-Plus use English numbered steps. LIBERO-GeoMotion retains the original Chinese Markdown expert notes, including recorded coordinates and replay results. Those notes also link to upstream CSV, HDF5, NPY, and combined videos that are not included in this package; the files distributed here are listed above.

🚀 Quick start

1. Explore a demonstration

Try RoboTwin: adjust a bottle, RoboDojo: fasten screws, or LIBERO-GeoMotion: move a blue bar onto a green pad. Open the corresponding videos/ directory to view the camera recordings.

2. Check the reference files

Install the tools following the GPT-GEAR installation guide. With FFmpeg and FFprobe available, run this example from the GPT-GEAR repository root. Replace /path/to/reference_demo with the absolute path to this package:

gpt-gear reference prepare --benchmark robotwin \
  --reference-video /path/to/reference_demo/RoboTwin/adjust_bottle/videos \
  --reference-text /path/to/reference_demo/RoboTwin/adjust_bottle/steps.txt \
  --trajectory /path/to/reference_demo/RoboTwin/adjust_bottle/trajectory.json \
  --reference-mode sequence --reference-fps 1 \
  --output outputs/robotwin_reference_check/reference

Use a new output directory. This command validates the files and builds a reference cache without starting a simulator or calling a model.

3. Add demonstrations to evaluation

The GPT-GEAR queue files configs/queues/robodojo.json, configs/queues/robotwin.json, and configs/queues/libero_geomotion.json already contain the task mappings below. Replace their environment paths and /path/to/reference_demo placeholders with your local paths before running.

For RoboDojo and LIBERO-GeoMotion with video, text, and full trajectories, increase the context window to 432000. Their queue files already set top-level "codex_context_window": 432000; retain it when copying a configuration. For a standalone gpt-gear run, add --codex-context-window 432000 explicitly, using a model that supports this window. The default 272000 can trigger context compaction before the full reference is inspected. The offline check above does not need this option.

Check the actual context window reported by the Codex session as well as the requested value. A custom model catalog can cap the window below the CLI setting and cause early compaction.

Run one RoboDojo task

Try fasten_screws with video, text, and trajectory references. Prepare the RoboDojo environment and model authentication following the GPT-GEAR guide, then activate the tool environment in the GPT-GEAR repository root. Replace every /path/to/... placeholder with your local path:

gpt-gear run fasten_screws outputs/robodojo_reference \
  --benchmark robodojo \
  --env-root /path/to/RoboDojo \
  --sim-python /path/to/robodojo/bin/python \
  --seed 0 --layout-id 9 \
  --codex-context-window 432000 \
  --reference-video /path/to/reference_demo/RoboDojo/fasten_screws/videos \
  --trajectory /path/to/reference_demo/RoboDojo/fasten_screws/trajectory.json \
  --reference-text /path/to/reference_demo/RoboDojo/fasten_screws/steps.txt \
  --reference-mode sequence --reference-fps 1

This command starts the simulator and calls the model. Use a new output directory; do not create it beforehand. The example uses seed 0 and layout 9, matching the RoboDojo evaluation queue. After the run, inspect the result with:

gpt-gear summarize outputs/robodojo_reference

Add references to a queue

For a custom queue, add a reference object to the matching episode. For example, the adjust_bottle episode uses:

{
  "reference": {
    "video": "/path/to/reference_demo/RoboTwin/adjust_bottle/videos",
    "text": "/path/to/reference_demo/RoboTwin/adjust_bottle/steps.txt",
    "trajectory": "/path/to/reference_demo/RoboTwin/adjust_bottle/trajectory.json",
    "mode": "sequence",
    "fps": 1
  }
}

This is an episode fragment; see the queue guide for complete configurations. Use the original files here, even if you have already prepared a cache. Absolute paths make configurations easier to move; relative paths resolve from the JSON configuration's directory.

You can supply text, video, or both, and optionally add trajectories alongside text or video. For input combinations and command-line examples, see the embodied ICL guide.

🗺️ Task mapping

RoboDojo

Task IDs match RoboDojo/<task>/ directory names. The GPT-GEAR queue contains 42 tasks and uses 34 same-name demonstrations. Its eight tasks without references are align_blocks, classify_objects_by_language, general_pickup, pick_from_conveyor_by_image, pour_by_language, solve_equation, stack_blocks_by_language, and store_tools_in_toolbox.

The additional spell_RoboDojo demonstration is included for exploration and is not part of that queue.

RoboTwin

All 50 tasks in the GPT-GEAR queue have a same-name demonstration at RoboTwin/<task>/. Each task uses the same demonstration for both clean and random configurations and all five evaluation seeds (0–4).

LIBERO-GeoMotion

One demonstration is shared across each listed task family. Scene IDs use the libero_rp/ prefix, and ranges below are inclusive.

Evaluation scenes	Task family	Demonstration directory
libero_rp/0–libero_rp/3	Blue bar onto green pad	task_000
libero_rp/4–libero_rp/7	Book upright in organizer	task_004
libero_rp/8–libero_rp/11	Bottle oriented in rack	task_008
libero_rp/24–libero_rp/27	Mango transfer around protected objects	task_026
libero_rp/28–libero_rp/33	Cream cheese into drawer, then close	task_028
libero_rp/34–libero_rp/35	Two mugs into cabinet, then close	task_034

Scenes libero_rp/12–libero_rp/23 have no demonstrations in this package; their queue entries omit reference inputs.

LIBERO-Plus

The 40 demonstrations use LIBERO-Plus/base_00/–base_39/, with one demonstration per base task:

Demonstration directories	Source suite	Count
base_00–base_09	libero_spatial	10
base_10–base_19	libero_object	10
base_20–base_29	libero_goal	10
base_30–base_39	libero_10	10

Each task directory follows the common layout above and includes agentview and robot0_eye_in_hand videos. Its trajectory.json uses the same field structure as LIBERO-GeoMotion: format, frames, fps, pose, action, sampling, and files. The files field records relative paths to the trajectory and videos.

The base_XX identifiers are demonstration IDs, not LIBERO-Plus runtime task indices. Match the base task name to a variant in your installed environment before choosing suite/task_id; for example, base_38 describes putting both moka pots on the stove and does not imply libero_10/8. Use --benchmark libero-plus and the selected task directory's videos/, steps.txt, and trajectory.json as the three reference inputs. The existing GPT-GEAR configs/queues/libero_plus.json does not attach these references automatically; add the matching episode's reference object using the pattern above.

💬 Join the community

We welcome feedback on the demonstrations, documentation improvements, and experiments with different reference inputs. Share your findings through GPT-GEAR Issues. For data or mapping questions, include the environment, task ID, relevant file path, and what you observed so others can reproduce it.

For contributions to the framework, see the contributing guide. Thank you to the RoboDojo, RoboTwin, LIBERO-GeoMotion, and LIBERO-Plus communities for their simulation environments and work on robotic manipulation!