Datasets:

---

ACERobotics

/

ACE-Data-0

Like 58

Follow

ACE Robotics 38

Hi ACE-Data-0 team,

Thank you for releasing ACE-Data-0! We are working with the table-scale tactile subset (revision 490ef48), and we have a few questions about the 256-channel glove format used in Table-scale/Tactile/ACE-Data-Tactile-1 … ACE-Data-Tactile-10 (raw/tactile/{left,right}/take-*.parquet, columns pressure_0 … pressure_255 and contact).

1. Channel layout / mapping to the hand-shaped grid

The paper (Sec. 4.2.4) mentions hand-shaped pressure grids remapped from the raw glove sensors, but we could not find the channel map in the release. Is it the same mapping as TouchAnything / EgoTouch (configs/pressure_position_mapping_{left,right}.json in the TouchAnything repo)?

For reference, here is what we observed:

All channels referenced by that mapping (138 left / 147 right) are active in ACE-Data-0, while the remaining channels carry only about 0.3% of the total signal.
Under that mapping, neighboring taxels are clearly correlated (mean r ≈ 0.38, vs ≈ 0.08 after shuffling the channel positions).

So it looks compatible, but we would like to confirm. If possible, could you share the official channel map, including finger assignment and orientation?

2. Bend / flex sensors

The EgoTouch glove has dedicated bend sensors (3 per finger) in the same 256-channel array. Does your glove also have bend sensors, and if so, which channels are they?

In the bend-sensor region of the TouchAnything layout, only one channel per finger reads continuously non-zero, even without contact:

left: pressure_209 / 212 / 215 / 218 / 221
right: pressure_34 / 37 / 40 / 43 / 46

The other bend-region channels stay near zero. Are those other channels unconnected, or do they serve another purpose?

3. Baseline correction and normalization

Many channels have non-zero offsets at rest. For example, on the left hand, the 5th percentile of the summed pressure-channel reading within 10-second windows stays above 100 counts in 456 of 494 takes.

The paper mentions calibrated normalization and baseline correction. Could you describe the procedure?

How is the baseline estimated: first frames, a rest period, per-take calibration, or a rolling baseline?
Is normalization global or per take/glove?
Is there any calibration to physical units?

Do you plan to release the processed pressure grids?

4. The contact column

What does contact represent, and how is it computed (glove firmware, a threshold on the pressure channels, or something else)? We found it hard to relate to the pressure readings:

It is 1 in about 64% (left) / 40% (right) of samples.
It switches about 85 times per minute (median across takes).
On the left hand, after excluding bend-like channels, its agreement with “max pressure channel > threshold” is at most about 0.82 across thresholds 0–50. The mean max pressure is about 30 counts when contact=1, vs about 21 when contact=0.

Is this field intended to be used as a contact label?

Thanks a lot for your help, and for the great dataset!