Danielberry0311/lingbot-va-franka-real-datasets

LingBot-VA Franka Real-Robot Datasets

This repository contains two real-robot Franka manipulation datasets prepared for post-training the currently open-source LingBot-VA framework (referred to here as LingBot-VA 1.0).

Although the archive and directory names end in _lingbot_va2, the contents use the LeRobot v2.1 + Wan2.2 VAE latent + action_config format consumed by the current open-source LingBot-VA training loader. The names are retained for compatibility with the local conversion records; they do not indicate the native LingBot-VA 2.0 semantic visual-action tokenizer format.

Files

Archive	Episodes	Robot frames	Size (bytes)	SHA256
place_objects_into_the_box_lingbot_va2.zip	274	288,103	6,177,440,409	d5019aeb79e684270d51c8cdfd4114743055633b3eb96777c964916f545034c2
stack_bowls_lingbot_va2.zip	308	277,620	5,329,059,153	855a9ec809da756449d912ba7b485ca980dd71fe0cb5762c18a20311e4c97e88

Each archive includes:

• LeRobot v2.1 metadata and parquet files;
• three synchronized camera streams;
• action_config segmentation and language annotations;
• precomputed Wan2.2 VAE video latents;
• UMT5 text embeddings and empty_emb.pt;
• conversion and validation metadata.

The two archives contain 582 episodes/action segments and 565,723 robot frames in total.

Recorded action semantics

The native action has eight dimensions:

x y z qx qy qz qw gripper

• EEF position and quaternion are absolute poses recorded from /franka_robot_state_broadcaster/current_pose.
• Quaternion order is xyzw.
• action[t] is the recorded state at t+1.
• Gripper convention is 0 = open, 1 = closed/grasp.
• Joint states are recorded directly from /joint_states and are retained in observation.state.
• EEF poses and joints were not reconstructed using FK or IK, and no RPY round trip was used.

Intentional preprocessing consists of nearest-timestamp synchronization, quaternion unit normalization/sign canonicalization, gripper binarization, the t+1 target shift, and training-time quantile normalization.

For training these absolute EEF actions with the LingBot-VA loader, keep:

env_type = "none"
action_per_frame = 12
used_action_channel_ids = [0, 1, 2, 3, 4, 5, 6, 28]

Do not use the RoboTwin relative-pose conversion.

Cameras

observation.images.cam_high        <- global_image
observation.images.cam_left_wrist  <- wrist_image
observation.images.cam_right_wrist <- right_image (third side-view stream)

The third key is retained for LingBot-VA loader compatibility and does not necessarily represent a physical right-wrist camera.

Download on a training server

Install the current Hugging Face CLI:

python -m pip install -U "huggingface_hub[hf_xet]"

Download the repository:

HF_XET_HIGH_PERFORMANCE=1 hf download \
  Danielberry0311/lingbot-va-franka-real-datasets \
  --repo-type dataset \
  --local-dir /data/datasets/lingbot-va-franka-archives

Verify the archives:

cd /data/datasets/lingbot-va-franka-archives

echo "d5019aeb79e684270d51c8cdfd4114743055633b3eb96777c964916f545034c2  place_objects_into_the_box_lingbot_va2.zip" | sha256sum -c -
echo "855a9ec809da756449d912ba7b485ca980dd71fe0cb5762c18a20311e4c97e88  stack_bowls_lingbot_va2.zip" | sha256sum -c -

Extract both datasets into one clean training root:

mkdir -p /data/datasets/lingbot_franka_train

unzip place_objects_into_the_box_lingbot_va2.zip \
  -d /data/datasets/lingbot_franka_train

unzip stack_bowls_lingbot_va2.zip \
  -d /data/datasets/lingbot_franka_train

cp \
  /data/datasets/lingbot_franka_train/place_objects_into_the_box_lingbot_va2/empty_emb.pt \
  /data/datasets/lingbot_franka_train/empty_emb.pt

Point the joint-training dataset_path to:

/data/datasets/lingbot_franka_train

Framework and model

• Training framework: https://github.com/robbyant/lingbot-va
• Base model: https://huggingface.co/robbyant/lingbot-va-base
• Project page: https://technology.robbyant.com/lingbot-va

License and responsible use

No standard open-source dataset license has been selected for this release (license: other). Public visibility makes the files downloadable but does not by itself grant unrestricted reuse rights. Contact the repository owner for use beyond the intended research and model-training collaboration.

Real-robot policies can produce unsafe actions. Any deployment must enforce workspace, velocity, acceleration, collision, timeout, and emergency-stop protections, and must preserve the recorded base frame, EEF/tool definition, camera ordering, quaternion convention, and gripper semantics.