dfmainsg/Physical_AI_TRumi-ds3

TRUMI DS3

This dataset contains robot manipulation demonstrations collected with TRumi (Trossen Robotics Universal Manipulation Interface) and processed with the TRumi dataset generation pipeline into a Zarr-formatted replay buffer for imitation-learning policy training.

Dataset Summary

TRumi is a handheld manipulation data collection system for Physical AI: a handheld parallel-jaw gripper with a wrist-mounted GoPro that records what the gripper sees and how it moves. Instead of teleoperating a robot, an operator performs the task directly with the handheld gripper, which makes it fast to collect large numbers of demonstrations across many objects, environments, and task variations without a full robot rig.

TRumi follows the Universal Manipulation Interface (UMI) approach: demonstrations are captured in end-effector space (gripper pose + gripper width) rather than a specific robot's joint space, which makes the data reusable across different robot embodiments before robot-specific training and validation.

Dataset Structure File Size Description dataset.zarr.zip ~55.7 MB The packaged dataset: a single Zarr archive containing all episodes in a flat, NumPy-backed replay buffer, with JpegXL-compressed camera images. dataset_plan.pkl ~153 KB The dataset plan produced by the pipeline's 06_generate_dataset_plan.py stage — a pickled record of which raw demonstrations were used to build the dataset (and which, if any, were dropped).

Per the TRumi dataset format documentation, the Zarr archive stores, per time step:

End-effector pose (3D position + axis-angle rotation)

Gripper width (how open the fingers are)

Demo start/end poses

Camera images (JpegXL-compressed)

Because this is a Zarr archive rather than a tabular file, the Hugging Face dataset viewer does not render it directly — load it with the zarr Python package (see below) to inspect its contents.

How This Dataset Was Collected

Data collection with TRumi has three phases:

Collect — An operator performs the task while holding the TRumi gripper. The wrist-mounted GoPro records video along with onboard IMU (accelerometer + gyroscope) data. A one-time mapping video of the workspace is also recorded so SLAM has a map to localize against.

Process — The dataset generation pipeline runs visual-inertial SLAM (ORB-SLAM3) to estimate the camera's 6-DoF trajectory through the mapped scene, fusing video with IMU data for robust, metric-scale motion. ArUco tags are used to calibrate gripper width. For each frame, the pipeline then extracts the end-effector pose and gripper width.

Output — The synchronized frames, poses, and gripper widths are packaged into this structured .zarr dataset (an .mcap per-episode variant is also supported by the pipeline), ready for downstream policy training.

Intended Uses

This dataset is intended for training manipulation policies via imitation learning (e.g. diffusion policy, ACT, or similar visuomotor policies) using UMI-style, end-effector-space demonstrations. Because the data is not tied to a specific robot's joint configuration, deploying a trained policy on a real robot arm requires an additional integration step — converting policy outputs into that robot's motion via inverse kinematics, interpolation, and on-hardware validation.

Limitations

Captured with a handheld gripper, not a robot arm — trajectories reflect human motion and may require smoothing/filtering before use on hardware. Pose estimates depend on the quality of the visual-inertial SLAM solve for each episode; episodes with poor SLAM tracking may have been dropped during dataset generation (see dataset_plan.pkl).