TIVA

Runtime temporal integrity for perception-driven robots.

Prithvi Bhat · prithvibhat19@gmail.com

The problem

Under peak load, perception lags while health checks stay green. Motion continues. Logs look fine. Failures show up as wrong picks, near misses, or days lost on intermittent floor issues.

What TIVA does

TIVA evaluates whether observations and perception outputs remain temporally valid before they influence downstream planning and actuation: whether an observation should enter inference, and whether published output should reach planners and actuators. Perception models remain unchanged; no retraining required.

Stage 1 today: temporal observation (JSONL traces), validation (KPI and classification), and governed replay on recorded perception workloads (videos below). Traces make overload visible; governance holds or skips when output is no longer contract-compliant. Next: live edge governor on the same contract.

Where it applies

How it fits

Two checks: before inference (admit or skip) and before actuation (forward, hold, or withhold). Timing health is derived from the temporal contract in runtime traces.

Sensors Gate 1 before inference admit / skip Inference unchanged Gate 2 before actuation forward / hold Planner Actuators TIVA temporal governance (capture → infer → emit)
Two governance gates on capture → infer → emit (shown in replay demos below).

Why not timestamps, QoS, or watchdogs alone?

Existing approachLimitation
Timestamp thresholdDetects age but provides no runtime governance
Dropping framesReduces backlog but does not govern published outputs
ROS 2 QoSManages communication quality, not temporal validity for actuation
WatchdogsDetect failures after they occur; do not govern degraded operation

Validation

Lab replay on recorded workloads demonstrates governance under stale perception: stale published output triggers a downstream false brake without TIVA; with governance, the stream holds until output is temporally valid (conveyor demo). Replay comparison shows wait-or-skip under stress on a real detector. Controlled lab conditions; mechanism proof, not field ROI.

Conveyor: stale position triggers a false brake (left) vs hold until output is temporally valid (right). Controlled synthetic demo.
Replay: ungoverned vs governed (TIVA). Older videos may show a prior internal label on the governed side.

Design partners

Teams seeing timing-related perception failures under load: short calls to understand the workload; recorded-trace evaluations later by mutual agreement.