Exploring Dex3 pressure sensing#

The pressure tools read Dex3 state, preserve raw samples and turn valid baseline-relative counts into a spatial display. They do not publish hand commands. The physical mapping study used the right hand marked 214-R-T.

Seeing the pressure display#

Taxel colors change at different locations on the rendered Dex3 hand. This is the RViz display; the physical contact producing the readings is outside the frame. Colors represent pressure-count changes, not calibrated forces. Silent, 11 seconds.

Download the tactile demo (MP4)

Follow a changing cell back to the named taxel mapping and the raw slot. A display can look plausible even when a slot is misplaced or its baseline was captured during contact; the retained raw samples let you investigate that.

Start with the raw message#

A taxel is one tactile sensing cell. A message slot is only a location in the raw array: many slots on this hand are unused. Establish which slots respond to touch before interpreting a colored hand model.

unitree_hg/msg/HandState contains nine pressure groups with 12 cells each: 108 raw pressure slots per hand. Each group also contains temperature values, lost and reserve fields. On the inspected right hand, 33 slots responded as taxels and 75 remained exactly at the unused value 30000.

Group ID	Active cell indices
0, 2, 4, 6, 7, 8	0, 2, 9, 11
1, 3, 5	3, 6, 8

An initial 45-sample, 4.87 s audit established the layout. A later 210-sample, 20 s untouched baseline confirmed that every inactive slot stayed at 30000. The map begins from an external diagram, but repeated live touches are the authority for this particular hand’s placement.

Topic rate changes what you can observe#

The source measured approximately 9.1 Hz on /lf/dex3/right/state and 750–770 Hz on /dex3/right/state in short runs. The lower rate was useful for guided mapping and RViz; fast taps or slip need the higher-rate stream. Recheck rates on the actual firmware and network rather than treating these historical measurements as specifications.

The tools subscribe to state and publish visualization markers plus optional joint states. They do not publish hand commands. After following the source workspace setup, a passive raw observation is:

ros2 topic echo --once --full-length /dex3/right/state unitree_hg/msg/HandState

The default visualizer uses the lower-rate topic:

ros2 launch dex3_pressure_tools dex3_pressure_visualizer.launch.py

If an existing launch already owns robot description and joint-state defaults, use the documented flags to avoid duplicate publishers. See the pinned README and lab protocols for the complete launch context.

Baseline and display model#

valid = finite(raw) and abs(raw - 30000) > 1000
baseline = median(valid untouched samples)
noise = p99(abs(untouched samples - baseline))
delta = max(raw - baseline, 0)

The untouched rerun had median/max standard deviation about 5.9/7.5 counts and median/max p99 absolute drift 16/24 counts. No active taxel drifted by 50 counts or more. That supported a 50-count visualization threshold on this hand under these conditions.

The free-touch session contained 1,274 samples over 139.7 s. Across active taxels, peak delta had median 2,280 and maximum 13,800 counts; 22/33 exceeded 500. A 500-count “definite touch” display threshold was conservative relative to those observations, not a force threshold in newtons.

Rebaseline while untouched. Do not let an existing contact become the zero reference. Preserve raw arrays and validity masks alongside the display; a color overlay alone cannot reveal an unused slot or baseline mistake.

Save an untouched baseline#

With the pressure package built and its ROS/message workspace sourced, the session recorder can record a baseline without a following touch session:

ros2 run dex3_pressure_tools dex3_pressure_session_recorder --ros-args \
  -p baseline_seconds:=20.0 \
  -p record_seconds:=0.0 \
  -p prefix:=dex3_pressure_baseline \
  -p output_dir:=/path/to/new-recording-directory

Keep the hand untouched and press Enter when prompted. The output includes JSON statistics and compressed NPZ samples with timing and motor positions. Check inactive slots, valid sample counts and per-taxel noise before using the result as a reference. The default topic is /lf/dex3/right/state; select state_topic explicitly when evaluating the higher-rate stream. This is a passive recording procedure; it does not close the hand.

Mapping the fingers#

Physical RViz checks established right-hand motor order as thumb 0/1/2, middle 0/1, index 0/1. The explicit 33-taxel YAML stores marker poses on named links. A misplaced visual taxel should be changed after repeated touch evidence, not simply because a mirrored diagram looks more plausible.

Later cube manipulation explored pressure as required grasp evidence and replaced that gate with opposed joint shortfall relative to measured empty close. That result does not invalidate the sensor study. It says pressure was not a reliable required signal for those cube contacts and placement. The contact-check explanation describes how achieved fingers constrain the carried route.

Evidence: dex3_pressure_tools README, observations, signal model and mapping documents at the recorded revision.

Follow the code#

Implementation concern	Code entry point	Responsibility
Raw recording	session_recorder.py · `Dex3PressureSessionRecorder._state_callback`	Store pressure arrays, receipt times and motor positions.
Message processing	pressure_visualizer.py · `Dex3PressureVisualizer._state_callback`	Extract slots, retain validity and compute positive baseline-relative deltas.
Untouched baseline	pressure_visualizer.py · `Dex3PressureVisualizer._finish_baseline`	Estimate per-slot median and noise; choose the display threshold.
Spatial display	pressure_visualizer.py · `Dex3PressureVisualizer._publish_timer_callback`	Publish mapped markers and optional visualization joint state.

The raw recorder retains information that a color overlay discards. The display depends on both a valid untouched baseline and a physically checked taxel map. Counts are not forces, and this passive path is separate from the later joint-based grasp-contact gate.

For whole-robot episode recording, pressure arrays in LeRobot, and dataset downloads, see Recording and LeRobot. This chapter covers how to interpret and visualize the hand’s sensor readings.

Checks and evidence to inspect#

Use the repository’s docs/OBSERVATIONS.md, docs/MATH_AND_SIGNALS.md and docs/MAPPING_AND_MARKERS.md at the recorded revision. The audit and touch results above are for the inspected right hand; no new tactile run was performed for the guide.