e0017 · e0017_merged · pick_up_bottle
One timeline, both views: what the arm did and what the gloves felt
A 22.63 s teleoperation episode — xArm6 carrying a BrainCo Revo2, driven by UDEX gloves through PICO. The simulated replay, the two cameras and the glove tactile are all indexed off the same 30 Hz grid, so one scrubber moves every pane to the same instant. The provenance notes at the bottom say what this episode is and is not good for.
Simulated replay against the cameras channel 1 vs channel 2
Simulationcalibrated FK
Scene cameratable, static
Wrist camerarobot flange
Glove tactile, laid on the hand 880 taxels per hand · 15 pads
Left glove
Right glove
Total pressure across the episodeeach hand to its own peak
Per-stream alignment onto the 30 Hz grid 680 frames
| stream | samples | measured Hz | aligned | hole frames | max gap ms |
|---|---|---|---|---|---|
| arm_right | 1150 | 29.92 | 98.4% | 11 | 25.1 |
| depth_head | 681 | 30.0 | 99.9% | 1 | 25.5 |
| hand_right | 1493 | 41.17 | 92.8% | 49 | 40.0 |
| imu | 4985 | 201.53 | 100.0% | 0 | 2.5 |
| pico | 3331 | 89.71 | 95.7% | 29 | 40.0 |
| rgb_head | 681 | 30.0 | 99.9% | 1 | 25.5 |
| tactile_left | 752 | 28.49 | 100.0% | 0 | 17.3 |
| tactile_right | 752 | 28.51 | 97.5% | 17 | 17.9 |
| wrist_left | 682 | 30.0 | 100.0% | 0 | 1.3 |
| wrist_right | 682 | 30.0 | 100.0% | 0 | 5.8 |
Revo2 fingertip pads raw only — not in the aligned artifact
| finger | status_ok | peak normal N |
|---|---|---|
| thumb | 6.5% | 0.000 |
| index | 87.4% | 0.070 |
| middle | 100.0% | 0.020 |
| ring | 100.0% | 0.510 |
| pinky | 0.0% | 4.050 · never reported |
Provenance. QC gate PASS — 0 fail, 8 warn, 44 pass. The operator marked the run a failure. Operator note: “didnt grab at roght locatipn” Judge training-worthiness from the gate and the stream table, not from this preview.
The gloves are on the operator, not the robot. The hand heatmaps are the human's side of the interaction (left peak 0.0168 N, right peak 0.4318 N); the pads that actually touch the object are the Revo2's, and those live only in raw/*/hand.jsonl — see the fingertip-pad table.
The cross-host clock offset is measured but approximate. Correlating arm motion against wrist-camera motion puts the offset at +8333.3 ms (correlation 0.438 against 0.245 at zero lag, resolution 440 ms), and the estimate is marked not confident — the peak is shallow. Camera frames in this page can therefore sit a few frames off the robot-state timeline.
The simulation pane is approximate in one known way.
The flange→hand mount rotation was set by matching the cameras
(180° about the tool axis — palm down, thumb on the base's +y side, verified
against the head and wrist cameras 2026-08-25), but the adapter plate's
translation is still unmeasured: the Revo2 base is drawn seated flat on the
flange face, so the plate's thickness is missing. The joint angles themselves are the
recorded ones, pushed through this machine's factory-calibrated kinematics.
rgb_head is a static table camera despite the name, and the wrist camera
is bolted to the robot flange.