jclarky/fr3_bowl_real

FR3 real-robot manual demonstrations

Real-world demonstrations on a Franka Research 3 (FR3) arm, collected to fine-tune an OpenVLA-OFT vision-language-action policy for Project #40, "A Large VLA Model for Industrial Robotics" (University of Auckland, 2026), with the VLAP4P codebase (scripts/fr3_demo_console.py, policy_core/fr3_demos.py). The same tasks exist in the project's Isaac Lab FR3 simulation; this dataset is the real-robot counterpart.

Dataset visibility: public.

Latest upload: bundles/2d7f5eac3f803938 (packaged bundle)

Task	Task id	Instruction	Episodes	Layouts	Minutes
1	fr3_bowl_on_plate	pick up the bowl and place it on the plate	10	10	6.8

Only episodes an operator reviewed in both camera views and accepted as successful task completions are included; stopped, failed and unreviewed attempts are not uploaded. Repeats of one layout (plan) are not independent: keep them in the same train/validation split.

All uploads in this repository: bundles/2d7f5eac3f803938.

How it was collected

• Not teleoperation and not policy rollouts. For each starting layout an operator photographed the scene, then taught joint-space waypoints by hand-guiding the arm (or planned paths in RViz), with gripper open/close steps. The robot then replayed the plan through MoveIt while it was recorded: every path checked for collisions and the workspace box, slowed to 0.08 m/s at the gripper and 0.8 rad/s per joint, and stopped above a contact-force limit (default 25 N). Objects were reset by hand to the starting photograph between episodes.
• Robot state: measured joint positions, TCP pose from the robot's kinematics in the fr3_link0 base frame (TCP fr3_hand_tcp), finger widths, external force/torque.
• Cameras: fixed overhead Azure Kinect (1280x720 colour) and a wrist-mounted RealSense D435i (640x480 colour). They are not hardware-synchronised: each 20 Hz sample takes the newest frame of each, and both cameras occasionally skip frames, so per-sample frame ages and the camera time difference are stored. Recording limits for this upload: camera frames at most 0.6 s old, cameras at most 0.6 s apart, samples at most 0.25 s apart, joints/force at most 0.1 s old.

Files

• bundles/<hash>/ (packaged): canonical/episode_<id>.h5 per episode, resampled to 20 Hz:
  • observations/overhead_rgb, observations/wrist_rgb: 256x256 RGB (overhead uses the policy's crop/rotation/correction recorded in each plan's provenance; wrist is a square resize)
  • observations/proprio: 8-D, TCP position (m), TCP orientation as axis-angle (rad), two finger joint positions (m)
  • actions: 7-D, achieved TCP translation / 0.5 m, achieved axis-angle rotation / 0.5 rad (both in the base frame), gripper intent (+1 open, -1 close). Labels come from measured motion, not from the commands sent.
  • timestamps, plus attributes instruction, task_id, camera-rig and calibration hashes, metadata.
  • recordings/ raw episodes (full-resolution JPEG frames, joints, force, per-sample timing), plans/ (starting photographs, taught waypoints, checked trajectories), BUNDLE.json (SHA-256 of every file).
• sessions/<name>/ (raw): session.json, plans/ and accepted raw recordings, for packaging elsewhere.

Use

With the VLAP4P checkout:

python scripts/fr3_demo_dataset.py package --session sessions/<name>     # raw sessions only
python scripts/fr3_demo_dataset.py verify --bundle bundles/<hash>
python scripts/fr3_demo_dataset.py export --bundle bundles/<hash> --out data/rlds --name fr3_real_manual

export writes an RLDS/TFDS dataset for OpenVLA-OFT fine-tuning (two images + proprioception, action chunk 8), registered by the repo's isaac_lab_vla.oft_entrypoint.

Notes

• Camera frames show the lab and may show people in the background.
• One real robot cell and one camera setup; the camera/calibration identity is recorded per episode and is distinct from the simulated rig. Do not mix it with simulated data under the same rig identity.