LeeHakHo/square_real100

Square (NutAssemblySquare) — 100 real Franka demonstrations

Real-robot counterpart to the square_d2 task in the ID90 / OOD auxiliary-pose study. A Franka Panda picks a square nut by its handle and drops it onto a square peg. Every episode is a successful insertion.

Two recording sessions on the earlgrey rig, 2026-09-07 and 2026-09-09, merged. It extends LeeHakHo/square_real60 with 40 further demonstrations rather than replacing it: square_real60 is the set the published checkpoints were trained on, and it is left untouched so those results stay reproducible. Use this one for new runs. Companion to the simulated datasets in LeeHakHo/mimicgen_aux_data.

Demonstrations	100
Samples	41,853
Control / recording rate	15 Hz
Episode length	315 – 1000 frames (21 – 67 s)
Size	1.10 GB
Format	robomimic HDF5

Layout

data/demo_<i>/
  obs/agentview_image            (T,84,84,3) uint8   Intel RealSense D435, static, scene
  obs/robot0_eye_in_hand_image   (T,84,84,3) uint8   ZED Mini, wrist-mounted
  obs/robot0_eef_pos             (T,3)   float64     panda_hand_tcp, robot base frame
  obs/robot0_eef_quat            (T,4)   float64     xyzw
  obs/robot0_gripper_qpos        (T,2)   float64     [+f, -f], robosuite convention
  obs/object                     (T,7)   float64     block 0 = nut [pos(3) | quat(4) xyzw]
  obs/aux_valid                  (T,)    uint8       see "Gating the auxiliary loss"
  actions                        (T,7)   float64     [dpos(3) | drot axis-angle(3) | grasp]
  rewards, dones                 (T,)                sparse: 1 on the final frame
  joint_qpos                     (T,7)   float64     arm joints, provenance
mask/train  (90 demos)   mask/valid  (10 demos)
Each demo group carries `source_session` and `source_traj` attributes naming the raw
recording it came from. demo_0..59 are the 09-07 session, demo_60..99 the 09-09 one.

Images were captured at 1280x720, centre-cropped to square and resized to 84x84. The crop keeps the whole workspace at this camera pose; squashing 16:9 straight to a square would have distorted every object.

The two sessions are interchangeable

The 40 new demonstrations were recorded two days after the first 60, so the thing to check is whether they are the same measurement. They are:

	09-07 (60 demos)	09-09 (40 demos)
Nut static before the grasp	5.20 mm median (2.9 – 19.7)	5.04 mm median (2.7 – 11.5)
Nut moved after release	8.04 mm median	7.82 mm median
Net travel	0.33 m (0.21 – 0.43)	0.31 m (0.22 – 0.39)
Frames jumping > 3 cm	2 per demo (1 – 3)	2 per demo (1 – 4)
FK cross-check at grasp	42.99 mm median	38.11 mm median
aux_valid	98.2%	98.2%
Gripper transitions	exactly 2 in every demo	exactly 2 in every demo

"Moved after release" is measured against the pose at the last closed-gripper frame, the same definition square_real60 reports, so its 9.8 cm worst case is the same demo and the same number here.

The camera did not move between the sessions. Measured under the shared extrinsics, the table plane agrees to 0.22° in tilt and 1.5 mm in height — the arm is bolted to that table, so its plane is a physical invariant and this is the reading that matters. A SIFT homography on frame 0 says 6 px, against a 0.6 px within-session baseline, but 6 px at fx = 912 is 0.4°, and the plane fit does not see it. One extrinsics covers both sessions.

Gating the auxiliary loss with obs/aux_valid

obs/aux_valid is 1 where the object pose in obs/object can be trusted and 0 where it cannot. If you train an auxiliary pose head on this dataset, multiply the per-frame auxiliary loss by it. 98.2% of frames are valid (770 of 41,853 masked out); per demo the valid fraction runs 95.5–99.4%.

Why the mask exists

obs/object is not ground truth. It comes from FoundationPose tracking RGB-D through each episode, and that tracking has one systematic failure mode on this task: the moment the gripper stops occluding the nut, the pose can snap. Measured across all 100 episodes, every one carries 1–4 frames whose pose moves more than 3 cm in a single 15 Hz step, and after the gripper opens the pose can drift up to 9.8 cm even though the nut is sitting still on the peg.

Those frames are not merely noisy, they are wrong, and a wrong auxiliary target is worse than a missing one: it pushes gradient into the shared visual encoder for an input that does not support the label. This is the same failure the simulated study handles for coffee_pod yaw with ROTATION_BLOCKS — supervising a quantity the image cannot determine turns the auxiliary head into a noise source rather than a regulariser.

How the mask was derived

From two physical facts about the task, with no hand labelling:

1. Between consecutive frames the nut cannot move 3 cm. At 15 Hz that is 0.45 m/s, far above the teleoperation speeds actually used (measured maximum end-effector step was 7–10 mm). A larger jump means the pose teleported, not the object.
2. The nut is static before the gripper first closes and after it finally opens. Within those windows, any deviation from the window's median position is measurement error; the threshold is 2 cm.

Frames within ±2 of a flagged frame are also marked invalid: a jump means the pose was already sliding on the way in. The exact procedure is aux_validity() in panda_utils/scripts/r2st_to_robomimic.py.

How to use it

Per-frame auxiliary loss, masked and renormalised by the number of valid frames rather than the batch size:

m = batch["obs"]["aux_valid"].float()                     # (B, T)
per_frame = ((pred - target) ** 2).mean(-1)               # (B, T)
aux_loss = (per_frame * m).sum() / m.sum().clamp(min=1)

Add aux_valid to observation.modalities.obs.low_dim in the training config so the data loader actually reads it — robomimic only loads the observation keys the config names, so without that line the key sits in the file unused.

Ignoring the mask is a valid first run. Training works without any of this; the auxiliary head simply receives 1.8% corrupted targets. Wire the gating in if the auxiliary arms underperform the baseline.

Two things that will bite you

There is no simulator behind this data. data.attrs["env_args"] records provenance only; the environment named there does not exist. Set experiment.rollout.enabled = false in the training config, or robomimic will try to construct a rollout environment from it and fail. This also means the study's eval_fixed_scenes protocol does not apply — real-robot evaluation is a separate problem.

Actions are reconstructed, not recorded. The demonstration logger (panda_utils/scripts/log_demonstration2.py, and the ROS 1 original it was ported from) builds a teleoperation action with input2action, sends it to the controller, and never stores it. Actions here are therefore derived from what the arm actually did:

dpos  = eef_pos[t+1] - eef_pos[t]              robot base frame
drot  = axis-angle of R[t].T @ R[t+1]          as OSC_POSE expects
grasp = +1 while the fingers are closed, -1 while open

For behaviour cloning that is arguably the better label — the policy should imitate the motion that happened, not a command the impedance controller only partly tracked — but it is not the original command, and the final step repeats the previous action because no t+1 exists for it.

Calibration and accuracy

Camera-to-robot extrinsics come from a ChArUco hand-eye calibration (panda_utils/scripts/r2st_handeye.py): the board's pose in the camera from PnP, its pose in the robot base frame from touching five board corners with the closed fingertips and reading forward kinematics. Residual 2.58 mm RMS over those five points; the fitted board plane is 0.67° from level and sits at z = +0.3 mm, consistent with the arm being bolted to that table.

Known residual error sources:

• D435 depth reads ~14 mm farther than RGB geometry on this white table at 0.86 m (measured against the ChArUco PnP solution over 7 corners, std 1.7 mm). Systematic, not noise. It does not affect the extrinsics, which use RGB only, but FoundationPose consumes depth, so object positions carry roughly that bias along the camera ray.
• Mesh proportions are approximate. The nut and peg meshes are parametric placeholders; scale_glb_to_pointcloud rescaled them 1.40x and 2.13x onto the observed point clouds. Physical measurements would improve pose accuracy.
• Forward-kinematics cross-check at the grasp frame puts the tracked nut 29–54 mm from panda_hand_tcp horizontally across the 100 episodes (median 41 mm). Most of that is geometric — the gripper holds the handle, while the tracked pose is the mesh origin — and its consistency suggests a systematic offset rather than random error.

The extra 40 add density, not coverage

Initial nut position in the robot base frame:

axis	60 demos (09-07)	40 demos (09-09)	all 100
x (toward/away from the robot)	13.6 cm range, 4.6 cm std	9.4 cm range, 3.1 cm std	13.8 cm range, 4.2 cm std
y (left/right)	21.2 cm range, 6.4 cm std	21.2 cm range, 5.8 cm std	22.4 cm range, 6.2 cm std
anisotropy (std)	1.39 : 1	1.88 : 1	1.47 : 1

The new session did not widen x. Its nuts sit in the far half of the older range (x 0.497–0.591 m against 0.453–0.590 m), so the combined x span grows by 2 mm while the anisotropy gets slightly worse, 1.39 : 1 to 1.47 : 1. For a ladder that grades generalisation by distance from the training distribution, the in-distribution window is therefore the same as square_real60's; what the extra demonstrations buy is sample density inside it, not a wider ID region. Spread x deliberately on the next session if the ladder needs more room.