jcapmeil/cupflow-data

CupFlow: real-robot cup rearrangement recordings

Real-robot recordings of a Franka FR3 rearranging six cups on a table, one cup per episode, from one layout shape to another (circle, rectangle, equilateral / isosceles / right triangle, scatter). The data was recorded between 2026-09-18 and 2026-10-04.

Size: 151 sessions, 906 episodes (6 per session), about 172 GB. Sessions are stored as one .tar each (about 1 to 2 GB) so the repo stays small in file count.

Start here

1. derived/episode_labels.jsonl has one row per episode: grasp type, wrist-measured pick position, commanded and measured placement, placement error in mm. Read derived/README.md first. It lists corrections that supersede the raw files (camera extrinsics, commanded vs measured placement, Vicon frame).
2. runs.jsonl / RUNS.md is the index of sessions.
3. Download one session to try it:

pip install -U huggingface_hub
hf auth login   # needed because this dataset is gated
hf download REPO_ID --repo-type dataset --include "sessions/circle_to_scatter_20261004_171010.tar" --local-dir cupflow
cd cupflow/sessions && tar -xf circle_to_scatter_20261004_171010.tar

Everything except the raw sessions (derived/, runs.jsonl, RUNS.md, command.csv, state.csv, scripts/) is stored unpacked and downloads on its own:

hf download REPO_ID --repo-type dataset --exclude "sessions/*" --local-dir cupflow

All sessions (about 172 GB): hf download REPO_ID --repo-type dataset --local-dir cupflow.

Layout

README.md  RUNS.md  runs.jsonl  command.csv  state.csv
derived/                         additive labels and calibration fixes (see derived/README.md)
scripts/                         verify_runs.py, vicon_runs.py, sync_out.py (collection-side helpers)
sessions/<from>_to_<to>_<YYYYMMDD>_<HHMMSS>.tar   one per session (151)

Inside a session tar:

<session>/
  session.json            session config: shapes, seed, workspace, z_release, vicon setup, ...
  run_session.json        run-level config and health
  final_positions.json    six cups' commanded final positions (robot_base, metres, [x, y])
  ep0001_iter000_cup3_place/    ... ep0006_...   one directory per episode

Session names read <source shape>_to_<target shape>_<timestamp>. Some sessions chain: the layout one session ends with is the layout the next starts from (the chain is checked in derived/layouts.jsonl).

Inside an episode directory (one cup picked up and placed):

File / dir	Content
meta.json	Episode metadata: instruction text, target slot, commanded move segments, grasp rung, health, stream stats with sha256, clock notes, code git commit
calibration.json	Camera intrinsics and extrinsics. base_T_cam is wrong for sessions from 2026-09-30 on (the ZED moved). Use derived/zed_refit_epoch1.json for those
state.jsonl	Robot state at about 50 Hz: joint positions q, velocities dq, torques T, dynamics terms M, G, C, the commanded action (joint_velocity), gripper fields, t_rel seconds
vicon.jsonl	Vicon pose of the arm end effector at about 30 Hz, in the Vicon world frame, quaternion order xyzw. Not registered to the robot base; apply derived/vicon_registration.json. Cups are not tracked
zed.svo2	ZED recording (SDK 5.0.7, HD1080, 30 fps, neural depth). Needs the Stereolabs ZED SDK to read
zed/depth_<i>_<ms>ms.png	ZED depth at 5 Hz, 16-bit PNG in millimetres, 0 = no reading, 960x540
realsense/frame_<i>_<ms>ms.png	Wrist RealSense RGB at about 30 Hz, 640x480
realsense/depth_<i>_<ms>ms.png	Wrist RealSense depth, 16-bit PNG (raw z16), 640x480
realsense/video.mp4	Video of the wrist camera. Present for only some episodes (498 of 906)
scene_{before,after}_{rgb,depth_mm,depth_view}.png	1920x1080 ZED scene snapshots taken with the arm stationary just before and after the episode. depth_mm is 16-bit millimetres; depth_view is a colourised preview

Filenames end in the time in milliseconds since the episode's t0, the same clock as t_rel in state.jsonl. meta.json has the clock offsets for the ZED and Vicon streams.

Coordinates: positions are in robot_base, metres, unless noted. One layout cell is 10 mm.

Top-level tables

• command.csv: commands received by the robot bridge (cmd0..cmd6), with wall and monotonic timestamps.
• state.csv: robot-side state stream (q, dq, tau).
• sender columns were replaced with REDACTED_IP; nothing else was changed.

Caveats

• released_xy / target_xy in the raw files are commanded. The cup lands a median 3.8 mm from that (p90 9 mm). Use placed_xy_wrist_measured in derived/episode_labels.jsonl where present (774 of 906 episodes).
• Camera extrinsics: use calibration.json for sessions before 2026-09-30 and derived/zed_refit_epoch1.json after.
• Vicon tracks the arm only, so there is no motion-capture ground truth for the cups.
• The raw sessions are never edited after recording; meta.json carries sha256s for the stream files.
• Run-level data is marked healthy or not in meta.json (health) and in runs.jsonl (healthy).

Access and license

This dataset is gated: request access on the dataset page and it will be approved manually. License and citation: to be added by the authors.

Contact: Jake Capmeil (via the Hugging Face profile that owns this dataset).