It now runs joy, teleop and dora-ulite6 with the probe declared and the
console shut, so a pad can be checked against the arm that will actually
move. ROBOT_IP in both arm dataflows points at the bench arm.
The run recipe drops dora up + dora start: dora 1.0.1 refuses to start a
dataflow holding git: nodes from a coordinator, reporting that no build was
run straight after one that succeeded. That costs the file dora topic, which
needs a coordinator to read the wire.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
An Xbox One S pad over Bluetooth numbers its controls differently from the
indices this node shipped with: the right stick is on axes 2 and 3, the
triggers on 4 and 5, and the buttons carry gaps at 2 and 5. Against the old
layout an untouched pad reported a resting trigger as full right-stick
deflection, which asked for 30 deg/s of pitch while lying on the desk, held
the lease for ever so GAMEPAD_IDLE_RELEASE_S never fired, and moved a real
arm into a controller error. X toggled freedrive and LB was the E-STOP, so
the resulting latch could not be released from the pad at all.
Layout now carries the swept indices. Each pad is also measured the first
time it is seen at rest: a stick must rest near centre and a trigger near
the bottom of its travel, and a pad that does not fit reads as absent and is
named in the log rather than driving an arm from a numbering nobody checked.
The measured rest becomes each axis's zero, so a worn stick's tilt is
subtracted rather than deadzoned away and a trigger that reads 0 until it
first moves stops reading as half pressed.
Rest is never re-measured while the pad is in use, because doing that under
a held stick would record the deflection as centre and the arm would move
when the operator let go. The probe drives from the same calibration, so
what it prints is still what the arm would get, and it shows the mismatch.
Tests name their controls through axes_for and buttons_for instead of
numbering them; the positional lists are why a wrong index passed a green
suite.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The browser reached this node through `dora-websocket-transceiver`: a second
node, a second port, and a round trip through the dataflow for every button
press, to carry a UI only this node can interpret. The node now serves the
page and holds the browsers' sockets itself on one port, so the console is
one URL and a consumer wires nothing for it.
`dora_teleop/server.py` runs aiohttp on a thread of its own and meets the
control loop at two non-blocking points: a bounded queue the tick drains, and
a snapshot handed back. Both queues are bounded deliberately — a flood is
dropped rather than grown, and a slow browser loses stale snapshots rather
than the newest — because nothing a browser does may stall the thread that
stops the arm. A port already in use stops the node rather than leaving it
announcing a console nobody can open.
The web deadman moved but did not change: it was a `clients` message from the
bridge and is now the open-socket count read each tick. One hop fewer, same
meaning.
Contract, so 0.2.0 -> 0.3.0: `ui_command` and `clients` are gone. `ui_state`
stays published for `dora topic echo` and the bench, with no browser reading
it. New WEB_HOST and WEB_PORT (7000; 0 opens no port).
The console itself is laid out like the arm's own pendant — two rings
quartered into arc sectors you hold, each with its out-of-plane axis as pills
above. Directions and signs match xArm Studio on the controller, `Y+` on the
left included: two consoles driving one arm that disagree about which way is
+Y is a trap, not a preference.
Also sends a velocity only when it changes, repeating an unchanged one on
JOG_KEEPALIVE_S (0.2 s) purely to beat the consumer's own expiry. Repeating
every tick made `dora-ulite6` re-plan the same motion twenty times a second.
That turned out not to be the whole cause of the stepping an operator felt —
the rest was fixed in dora-ulite6 v0.3.1, pinned here — but sending 18
messages where 83 went before is worth keeping on its own.
`dora-joy` moves to v0.1.1: v0.1 is an annotated tag, and dora 1.0.1 records
the tag object's id as the commit hash and then rejects its own checkout for
sitting on the peeled commit. Neither dataflow could be built before this.
`twist` and `command` are the ports a consumer should wire from now on, and
`jog` and `request` are deprecated. Nothing breaks yet — both still publish —
but the version should say the contract moved.
Both pull sibling nodes by `git:`, but dora resolves the wiring without
fetching them, so this stays offline. It is what catches a drift between the
robot dataflow's output_types block and the tag it pins.
Also ignore `git/`, where dora build leaves those clones.
The page was a status readout: it sent three of the nine commands the node
serves, so the browser could take the lease and stop the arm but not move it.
It now holds a direction to move and releases to stop, with pads for
translation and rotation, gripper, freedrive, home and a speed slider. It sends
a fraction of full deflection rather than a speed, so it never learns the
robot's units — the same vector a stick produces.
Plain HTML and no build step: the bridge serves the directory as it stands, and
a console that cannot be fixed without a toolchain is the wrong thing to have
between a person and a robot.
Details that matter on a machine that moves: the pointer is captured so sliding
off a button still releases it; releasing sends jog_stop rather than relying on
the deadman; blur, tab-hide and a closed socket release everything; and the pad
is disabled unless the console holds the lease, so a button that would be
refused looks refused.
test_console.py checks the page against ui.py both ways — every message type it
sends is one the node serves, every telemetry key it reads is one the node
sends. Neither language catches a rename.
The pad moves to dora-joy, so this node keeps no hardware dependency and the
pad becomes reusable. It arrives on `joy` as a plain reading; `JoyInput.read`
treats silence past JOY_TTL_S as absence, because that node dying looks like
silence rather than like an unplugged pad and the arm cannot tell them apart.
Both control sources are now normalised to a [-1, 1] six-vector before anything
is scaled, and the scaling happens in one function. The web path used to pass a
velocity straight through without going near mapping, which is why the console
had to speak the arm's units. It no longer does.
`twist` (SI) and `command` (robot-agnostic verbs) are the standard ports. `jog`
and `request` still carry the same intent in dora-ulite6's units and service
names so the two repositories can move independently; LEGACY_REQUEST is the
only place that knows what that node calls a verb. Wire one or the other, never
both. `joint_jog` is specified in the manifest and deliberately not built.
Schemas are declared rather than inferred: pyarrow reads a null `value` as its
own type, so an inferred `command` would change its port's type between one
verb and the next.
Three bugs found while moving this code, each with a test:
- `Sample.active` compared raw axes against a threshold, but SDL rests a
trigger at -1.0, so every connected pad read as in use for ever: the pad took
the lease on plug-in and GAMEPAD_IDLE_RELEASE_S never fired.
- The e-stop latch was never consulted when deciding a velocity, so Back
stopped the arm for one tick and the next stick nudge drove it again. Nothing
on the pad could clear it either. It is now checked above the lease, and X
releases it and clears the controller's error, as the console's RESET does.
- Buttons were read before the arbiter granted the lease, so the first press
after the pad had been idle was swallowed: not owned yet when read, and no
longer a rising edge by the next tick.
The dataflow turned on `enable_debug_inspection` without the README saying why
anyone would want it, which made it a line of configuration nobody could act on.
Documents the three commands it buys — topic list, topic echo, topic hz — and
that they need a coordinator rather than `dora run`. hz is the one that checks
the timings this node promises, so it names the measurement to expect.
`dora topic echo` and `dora topic hz` only read the wire when the dataflow asks
for it. Turning it on here means the velocities this node sends can be watched
live, and their cadence measured, without the bench printing anything and
without a robot — which is most of what testing this node with only a gamepad
needs.
Measured with it: the get_state poll runs at 1.88 Hz against a configured
STATE_POLL_S of 0.5.
Two problems, both found by the validate job once it could finally run.
categories named `teleop` and `input`, neither of which dora defines. The set is
fixed — sensor, actuator, robot, transform, filter, ml-inference, llm, speech,
communication, recorder, visualization, simulator, debug — and an unknown
variant makes the whole manifest unparseable, so every type annotation in it was
being ignored.
dataflow.yml spanned three repositories through sibling paths, which meant dora
could not resolve the other nodes' manifests and reported their typed outputs as
unannotated. It also could not be validated in CI, which clones one repo. The
house pattern is a self-contained dataflow, and this now follows it: teleop plus
a bench stand-in for the arm and the browser, both from this checkout. The real
three-node wiring moves to README.md, where a consumer would look for it.
The bench answers every service, publishes plausible telemetry, and prints the
velocities it is sent, so the node can be driven from a real gamepad with no
robot in the room. It is not a simulator and must not become one: it never
integrates a velocity or moves a pose.
The axis and button indices in mapping.py are SDL's documented layout rather
than an observed one, and until now the only way to find a pad that numbers
things differently was to watch a real arm move the wrong way.
`dora-teleop-probe` prints the raw axes, the buttons it sees and the velocity
that would be sent, live, with no dataflow and no arm — it cannot move anything.
It goes through the same Gamepad and velocity() the node runs, so what it prints
is what the arm would get; that is the whole reason it is worth having, and the
reason it must not grow its own copy of the mapping.
Every axis is listed, including ones the layout does not use, because an axis
marked unused that moves when a stick is pushed is exactly the finding this is
for. The centred bar makes a worn stick that does not return to zero obvious.
dora-node.yml is the typed contract a consumer wires against, and the only one
they get: no source package is shared between node repositories. The jog,
request and ui ports stay untyped, because dora treats a port carrying
request_id as pattern-correlated and several Arrow schemas legitimately flow
over each — one per service, one per console message.
dataflow.yml runs the whole operator stack from checkouts: this node, the
WebSocket bridge and the arm.
web/ holds a placeholder page, not the real console. It exists so the dataflow
has something to serve and so the node can be checked end to end before the
Control section is ported across from the monolithic repo. It reads telemetry
and can take, release and stop; nothing else.
The loop owns the one decision this node exists to make: whether a velocity
reaches the arm this tick, and from whom.
An arm under a velocity command keeps moving until told otherwise, so the
question is never whether to start but whether anything has happened that means
it must stop. Every path through _wanted_velocity that is not an owner actively
asking for motion returns a stop, so a case nobody thought of fails by halting
rather than by carrying on. The default has to be still.
Three ways a source goes away, all ending in the same place. It goes quiet, and
the watchdog fires after JOG_TTL_S. It goes away outright — the pad unplugged,
the last browser gone — and the arm stops at once, without waiting out the
timer, because an unplugged pad is not a centred pad. Or the process itself
goes away, and one all-zero velocity is sent before it does: a motion lives in
the controller, not here.
The stop is sent once, not every tick. A stream of zeroes would keep the arm
node believing a jog is live and hold robot_state at moving forever.
E-STOP is not gated on the lease, from either source. Somebody reaching for the
stop is the one moment when asking who owns the arm is the wrong question.
The console's telemetry keys are a contract: app.js reads arm_state, owner,
estop, vacuum, freedrive, position and joints by name, so renaming one breaks
the page silently. They are reproduced rather than improved for that reason,
vacuum included, which is the console's name for whatever tool is bolted on.
The two vocabularies do not line up and this is where they are reconciled. The
arm node reports a driver's lifecycle; the console draws an operator's screen.
The ordering in arm_state is the ordering of what an operator most needs to
know: a latched e-stop outranks everything, because the arm may well read ready
and saying so would invite somebody to push it.
A message this node does not recognise is dropped rather than refused. The
console and the node version independently, and a page from a newer build
saying something we have not learned yet must not take the arm down.
Two sources want one arm and the arm cannot take both. Putting the decision in
one place means there is exactly one answer to "why did it not move" rather than
one answer per input path.
The model is ported from the monolithic repo's RobotService lease rather than
replaced with last-writer-wins, because the console is already built on it: its
jog pad is gated on owning the arm and it draws an owner chip. A simpler scheme
would mean rewriting the page to match the node instead of the other way round.
The gamepad takes the arm implicitly, on the first nudge of a stick, so nobody
has to reach for a browser to use the thing in their hands — and gives it back
after sitting untouched, so a pad left on the bench cannot lock the browser out
forever. The web takes and releases explicitly, and may preempt: an operator who
can see the arm is better placed than this code to decide they should have it.
Pure functions, no state and no hardware, because this is the part that has to
be right and the cheapest to check.
Analog and blended: how fast the arm moves depends on how far the stick is
pushed, and several axes move at once. That is the whole reason a pad beats the
web console's four-sector dial pad, which can only ask for one axis at a time at
one fixed speed.
The deadzone rescales what is left of the travel rather than just clipping it,
so the arm starts from zero just outside the band instead of jumping to a
fraction of full speed the instant the stick leaves it. Expo bends the response
because fine positioning happens near centre, where a linear stick gives almost
no resolution.
Up on the stick sends the arm along +X, matching the console's dial, so the two
agree about which way is which.
The reader owns the pad on a thread of its own, and nothing outside that file
may touch pygame: everything else sees a plain immutable Sample. That is what
lets the rest of the node be tested with no hardware, and what keeps SDL's
quirks out of the arm logic.
Sample.present is the field that matters. A pad that has been unplugged reads as
absent rather than as centred, so "the operator let go" and "the pad fell off
the desk" stay distinguishable — they call for the same stop, but only one of
them is worth logging.
Only the display and joystick subsystems are initialised. pygame.init() also
opens audio, which on a headless robot fills stderr with ALSA errors for a
device this node never uses; the display is needed despite there being no
window, because SDL refreshes joystick state during the event pump.
Settings are parsed once at startup so a deadzone that would silently make the
pad inert stops the node before it joins the dataflow.
Ignore rules, markdown lint configuration and the pointer to AGENTS.md, so the
node itself arrives as a reviewable diff rather than as an initial import.
cristhianandClaude Opus 5