Files
PX4-Autopilot/docs/en
Saibernard 4da3cff386 fix(mc_nn_control): guard the model, the observations and the output (#28504)
* fix(mc_nn_control): reject a model the module cannot feed

The documentation tells users to swap the model array in control_net.cpp,
and the module then wrote 15 floats into whatever input tensor came back
and read 4 outputs from whatever output tensor came back, with no check
of either. A model with a different input width would write past the
tensor in the arena with no diagnostic. Reported in #28417.

Check the schema version before the interpreter is built on the model,
then the number of input and output tensors and the element count and
type of each, and refuse to start with the reason in the log when they
do not match. The checks live in a header without module dependencies so
they can be unit tested, and the operator registration moves to a header
so a test can load the shipped model through the real interpreter.

Assisted-by: Claude:claude-fable-5
Signed-off-by: Saibernard Yogendran <bernie97@seas.upenn.edu>

* fix(mc_nn_control): guard the observations and the network output

The network was fed whatever the position and attitude topics last held,
with no check of the validity flags, the age or finiteness, and its
output went to the motors with any non finite channel mapped to a stopped
motor while the other three kept thrusting. Reported in #28417.

Check the observations on every cycle, in the mode or not: the position
validity flags, the age of position, attitude and angular velocity
against the same limits the raptor module uses, and finiteness. The
network is not run on observations that fail, and a non finite output is
not sent. In both cases the last command of this session is repeated,
which is what the mixer would apply anyway, and the arming check reply
says the mode cannot run, so the commander leaves it through its own
fallback within one of its external check intervals. The repeat keeps
the actuator topic and the lockstep simulators going, which wait for an
output on every step. The first fault of an outage is reported once. A
non finite output is reported once and the network is not trusted again
until the mode has been left.

The module runs on angular velocity updates, so a lost angular velocity
stream would never have woken it. It is now a scheduled work item that
re-arms a delayed run on every cycle, which keeps it registering,
checking and answering the commander when the updates stop.

The trajectory setpoint starts unset and is reset to the current
position when the mode is entered without one. Before, the member was
never initialized and the reset only happened when a position update
landed on the same cycle as the first angular velocity update.

In SIH, a GPS loss while flying the neural mode produced the fault event
the same instant the position validity dropped, inference stopped, and
the commander was in Descend within its next check. A gyro loss on a
real time build produced the fault event 20 ms after the last sample and
the commander was in failsafe 150 ms later.

Assisted-by: Claude:claude-fable-5
Signed-off-by: Saibernard Yogendran <bernie97@seas.upenn.edu>

* test(mc_nn_control): cover the model, observation and output checks

Eleven cases on the checks the module runs before trusting its model,
its observations and its output: every position validity flag, each
staleness limit at its boundary and pinned to the raptor module's
values, never set and future stamps, non finite values in each
observation, the order in which faults are reported, every channel of
the output, and every way a model can fail the layout check.

One more loads the shipped model through the real interpreter and runs
the module's layout check on it, so a swapped model array the module
cannot feed fails here before it fails on a vehicle.

Assisted-by: Claude:claude-fable-5
Signed-off-by: Saibernard Yogendran <bernie97@seas.upenn.edu>

* fix(mc_nn_control): start the observation caches empty

The three cached observations were plain members, so the first
observation check read whatever the memory held before any of the
topics had published, and a stream that had not arrived could pass
that first check by chance.

They start empty now, which the first check reports as an invalid
position. A functional test builds the module in storage filled
beforehand and runs the check with nothing published.

Assisted-by: Claude:claude-fable-5-1
Signed-off-by: Saibernard Yogendran <bernie97@seas.upenn.edu>

* ci(tests): run every mc_nn_control test in tests_neural

The default filter of the target only matched the first test of the
module. It now selects every test of the module, and it is quoted so
the shell does not read the alternation as a pipe.

Assisted-by: Claude:claude-fable-5-1
Signed-off-by: Saibernard Yogendran <bernie97@seas.upenn.edu>

---------

Signed-off-by: Saibernard Yogendran <bernie97@seas.upenn.edu>
2026-10-03 11:27:21 -06:00
..
…
…