Agent authorizes. Jev reviews anomalies. Local guards constrain motion.
The Agent proposes bounded actions and an authorized candidate set. Jev reviews execution anomalies; versioned evidence and local checks govern action application. Scene and RGB views are simulator assets. Bird's-eye context is for visualization only; robot artwork is illustrative.
Project video
AgenticJevNav in four minutes.
Background, motivation, method, simulation results, and physical robot case studies. English narration and captions · 4 min · 1080p.
01 / Overview
Keep decisions aligned with execution.
Natural-language object search requires target identification, exploration, and motion from incomplete visual information. Execution anomalies, goal revisions, and remote inference latency can lead to ineffective actions, stale decisions, and incorrect termination. AgenticJevNav coordinates Agent planning, anomaly-triggered Jev verification, and local motion constraints without navigation-specific weight updates.
The Agent proposes bounded actions and an authorized candidate set from RGB, odometry, instance-evidence memory, and execution feedback. Jev reviews structured evidence after blocked or failed execution, or two consecutive low-progress segments. Mission, revision, skill, snapshot, freshness, pose, and once-only checks reject obsolete results before a single executor applies them.
Across six simulated task families, full A2 achieves 73.33% SR (44/60), 21.67 percentage points above the next-best Agent-only aggregate control under differing historical protocols. A GENISOM L1 quadruped deployment achieves 70.0% task success (21/30) across six indoor task families, including timed search reporting. Traces show veto-triggered review and online goal revision; grounding, approach planning, and runtime reliability remain limitations.
02 / Method
Reason globally. Verify on events. Constrain every action.
A
Anomaly-triggered review
The Agent retains goal and planning authority. Jev reviews blocked or failed execution and two consecutive low-progress segments, selecting only from the Agent's authorized candidates. Valid arrival claims use a direct-stop path; Jev does not certify task success.
B
Versioned evidence consumption
Mission, revision, skill, and snapshot identify each decision. Freshness, pose, authorization, and once-only consumption checks govern application. Goal corrections revoke old requests and motion; temporal validity does not establish target identity.
C
Bounded local execution
A single executor enforces leases, command TTL, and local RGB stop/veto. The evaluated loop is serial: motion ends before the next planning image, and the robot waits while decisions are pending. The nominal 6 Hz monitor rate is not a measured stopping latency or a physical safety guarantee.
GIFs play at 6× speed. The 3D view reconstructs recorded simulation motion. T6 success follows the cohort's timed-report rule and does not establish exhaustive search coverage.
Full A2 runs on a GENISOM L1 quadruped with Jetson Orin NX 16 GB navigation compute, RK3588 motion control, front RGB, onboard odometry, remote Agent inference, and event-triggered Jev. The policy receives no depth, LiDAR, maps, or target coordinates. A separate scene camera records external context.
70.0%Task SR / 21 of 30 trials
6Task families / five trials each
60Scene-camera and first-person recordings
Experimental setting from the paper. Object roles depend on the task instruction; the image shows the setup rather than trial outcomes.
Real-robot results
Task success by real-robot task, five trials per task
Task
SR
Trials
R1 Find a chair
60%
3 / 5
R2 Select the larger bin
40%
2 / 5
R3 Chair beside the table
80%
4 / 5
R4 Detour around a bin
80%
4 / 5
R5 Chair-to-suitcase revision
60%
3 / 5
R6 Timed non-detection report
100%
5 / 5
Trials collected November 9–11, 2026 in one indoor setting. R1–R6 are the physical counterparts of the six simulation task families. R6 uses a 200 s search cutoff and evaluates reporting within the searched scope.
Mean session duration is 196.3 s, with 282 total Agent + Jev requests across 30 trials (9.4 per trial). Simulation and physical execution use platform-specific motion bounds and TTLs. Traces record veto-triggered Jev review and chair-to-suitcase goal revisions. Two collisions, two wrong-goal arrivals, four timeouts, and one visual-monitor runtime failure delimit current capability. Arrival-policy changes during collection make this a descriptive deployment series; no matched physical control establishes a causal Jev benefit.