From scene state to safety evidence, in one deterministic path.
FieldSpace converts tracks, ego state, map context, and route intent into a deterministic risk field. In observer mode it emits replayable warnings and audit traces; in benchmark mode it can generate bounded trajectory candidates for closed-loop evaluation.
Three layers, observer budget under 50 ms
Every layer has a typed contract and a deadline. If any layer fails its contract, the observer emits a traceable fallback recommendation for review.
Route-relative risk
Projects scene actors and route context into the local risk field. In benchmark mode, this can also produce a trajectory candidate for closed-loop comparison.
- lookahead: 4-12 m speed-adaptive
- max_steer_rad: 0.52 (~30°)
- max_steer_rate_rad_s: 0.5
TTC and risk thresholds
Risk timing comes from explicit thresholds and scene-derived quantities. The output is replayable, auditable, and comparable against partner or benchmark baselines.
- kp / ki / kd: 0.4 / 0.05 / 0.01
- target cruise: 8.33 m/s (30 km/h)
- max_accel / decel: 2.0 / 5.0 m/s²
Fallback recommendation
Finite-state machine that monitors upstream contracts. On violation it records the recommended fallback phase and trigger reason rather than emitting an opaque alert.
- comfort_decel: 2.5 m/s²
- emergency_decel: 5.0 m/s² (TTC < 2 s)
- shoulder_drift: toggleable
The part reviewers ask about first
Safety reviewers want to know what happens when inputs degrade, an ODD boundary is crossed, or collision risk rises. The observer keeps those trigger classes explicit and replayable.
pub enum MrmPhase {
Idle,
Decelerate, // comfort ramp to near-stop
HoldLane, // maintain lane at low speed
DriftToShoulder, // if shoulder is mapped
Parked, // wheels stopped, brake held
Failed, // timeout escalation
}
// Canonical demo transcript:
// t=0.00s phase=Decelerate 12.00 m/s
// t=2.40s phase=DriftToShoulder 1.00 m/s
// t=8.95s phase=Parked 0.00 m/s
// [mrm] stopped after 8.95sReviewable because it's bounded
FieldSpace is not a certification by itself. It produces the bounded, replayable artifacts that make a safety review more concrete: inputs, thresholds, risk fields, trigger reasons, and output traces.
Functional Safety
Supports decomposition review by keeping the observer path deterministic, scoped, and traceable.
SOTIF
Helps structure edge-case review with explicit ODD assumptions, thresholds, and replayable scenarios.
Fallback Behavior
Provides documented trigger classes and fallback-state traces for technical review.
Safety evidence as a replay, not a black box.
Walk through MRM transitions with us against your own incident logs. You pick the scenario; we replay it deterministically.