Is this page helpful?

Perception Container Monitor (PCM)#

The Perception Container Monitor (PCM, nv_perception_monitor) assesses whether the AI perception pipeline can be trusted. It reports a single pipeline verdict, AI_PIPELINE_VALID or AI_PIPELINE_INVALID, to the SEI daemon. SEI forwards the verdict through the safety decision path, so PCM is an input to the SDM safety decision.

Operating modes#

PCM has two modes, selected with --mode LEARN|ACTIVE. The default is ACTIVE.

PCM Operating Modes#

Mode

Description

LEARN

Runs only the Data Validation Monitor (DVM). It observes mdx-frames for --learn-duration seconds (default: 300), writes per-sensor DVM baselines, then exits. It does not register with SEI or report pipeline verdicts.

ACTIVE

Registers with SEI, sends heartbeats, and combines container-health and DVM results into debounced pipeline verdicts. It reads the baselines written by LEARN.

What PCM checks#

In ACTIVE mode, PCM combines two checks:

  • Container health — Checks only the exact container names listed in pcm_monitor.cfg through the Docker Engine API. A missing container, lifecycle fault, unhealthy health check, or unreachable Docker socket is a fault.

  • Data Validation Monitor (DVM) — Consumes mdx-frames through its own Kafka consumer group. It validates the learned sensor set, object plausibility, frame rate, Kafka connectivity, and decodable frame data.

PCM evaluates each check on a one-second cadence. Any enabled check that is bad makes the combined result bad. PCM emits AI_PIPELINE_INVALID after two consecutive bad evaluations and AI_PIPELINE_VALID after four consecutive good evaluations. Verdicts are emitted only when this state changes.

Configuration and baselines#

PCM reads its ACTIVE-mode monitor configuration from /opt/nvidia/psf/bin/pcm_monitor.cfg by default. Set PCM_MONITOR_CONFIG to select another file. The configuration can enable or disable container_health and data_validation; at least one check must remain enabled.

LEARN writes one pcm_<sensor>_baseline.cfg file for each observed sensor. PCM resolves the baseline location in this order:

  1. --baseline-config <file>

  2. PCM_BASELINE_CONFIG

  3. /opt/nvidia/psf/bin/pcm_baseline.cfg

The selected path identifies the directory that holds the per-sensor baseline files; it is not a consolidated baseline file. LEARN writes pcm_<sensor>_baseline.cfg files to that directory. ACTIVE scans the same directory and loads every matching file, so no consolidation or additional file needs to be supplied after LEARN. Use the same baseline location for LEARN and ACTIVE.

For a remote perception stack, disable the container-health check and pass the remote Kafka endpoint with --kafka-broker <host:port>; otherwise containers that are not local appear absent and cause a fault.

Launcher behavior#

launch_hoisa.sh starts PCM by default in learn and active modes. It does not start PCM in skip mode. Pass --skip-pcm to suppress PCM in learn or active mode. Its launcher log is /var/log/psf/pcm.log.

See also