Runtime assurance for autonomous robots

Trust the planner.
Verify the command.

A deterministic safety layer that predicts uncertain collisions, overrides unsafe black-box planner commands, and explains the intervention with counterfactual evidence.

Black-box planner command
Uncertain reachable-tube prediction
Chance-constrained risk envelope
Reject + search
Allow unchanged
Applied command + auditable evidence
Reproduced warehouse-crossing result

Unsafe intent becomes measurable evidence.

An autonomous mobile robot and an uncertain forklift enter the same blind crossing. Aegis evaluates the original command plus 25 bounded alternatives, then applies the least-deviating candidate inside the configured risk envelope.

Commanded risk
99.526%
Applied risk
0.000311%
Min clearance
1.39 m
p95 latency
6.75 ms
Aegis runtime-assurance evidence summary
Engineering depth

A safety boundary that exposes its assumptions.

01 / UNCERTAINTY

Time-growing prediction

Obstacle position covariance expands with acceleration uncertainty instead of assuming perfect future motion.

02 / SIMPLEX

Planner-independent guard

The planner remains replaceable and untrusted; the supervisor owns the final command decision.

03 / EXPLAINABILITY

Counterfactual evidence

Every intervention records what changed, when risk appeared, and how much risk the alternative removed.

04 / DETERMINISM

Repeatable replay

Scenario input, control lattice, decision JSON, and report generation are deterministic and tested.

05 / FAILURE SEMANTICS

No false safety claim

When no candidate meets the budget, Aegis labels the result minimum-risk rather than calling it safe.

06 / EXTENSIBILITY

Research path included

ROS 2 adapters, control barrier functions, calibrated prediction, and HIL braking validation are tracked publicly.

Scope and integrity

Prototype claims stay testable.

Aegis is a research prototype, not a certified safety component. Measurements were reproduced on one arm64 machine and are not hard real-time guarantees. The public build log records the AI-assisted development window exactly and makes no unsupported comparison with other engineers.