Agentic navigation must maintain goals and adapt actions from incomplete evidence despite execution failures and delayed inference. We present AgenticJevNav, which links goal-directed planning, evidence memory, event-triggered verification, and bounded execution. The Agent plans from RGB, odometry, memory, and feedback, retaining goal ownership while authorizing a bounded action set. Jev reviews structured evidence after blocked/failed execution or two consecutive low-progress segments. Version, freshness, and once-only checks govern decision application. A single executor and independent RGB reflex constrain motion. Goal revisions revoke obsolete requests and actions before replanning. Across six simulated task families, AgenticJevNav achieves 73.33% SR over 60 trials, 21.67 percentage points above the next-best Agent-only aggregate control under differing historical protocols. A real-world GENISOM L1 quadruped deployment achieves 70.0% success over 30 trials, including bounded search reporting. Together, the simulation and physical-robot results demonstrate the effectiveness of the proposed architecture, while remaining failures highlight opportunities to improve grounding, approach planning, and runtime.
Figures & tables
Fig. 2: AgenticJevNav architecture. (a) Scene context and RGB examples from separate episodes. (b) Versioned evidence with uncertainty, timestamps, and capture poses. (c) Agent planning from RGB, odometry, memory, and feedback. (d) Direct proposal adoption or anomaly-triggered Jev review. (e) Shared version, pose, age, authorization, and once-only checks. (f) Independent CPU NanoDet-based RGB stop/veto. (g) Bounded execution with leases, command TTL, and feedback. Arrival claims use direct stop. The policy receives no maps, depth, LiDAR, or target coordinates. The navmesh supports only simulated constraints and scoring.
Method
Venue
SR (%) ↑
NE (m) ↓
Path (m) ↓
Time (s) ↓
Wait (%) ↓
Calls ↓
Navmesh constraint (%) ↓
SPL ↑
Agent controls
A0: Agent-only
—
51.67
0.6112.9%
2.49
120.40
69.19
9.20
30.00
0.469
A1: Agent + rule events
—
40.00
0.84
2.24
106.85
67.95
8.78
18.33
0.474
A2: AgenticJevNav (ours)
Ours
73.3341.9%
0.72
3.18
140.55
79.40
12.42
15.000.0%
0.59124.7%
Adapted navigation methods
VLFM [ 3 ]
ICRA’24
0.00
2.35
5.16
N/A
78.74
71.52
43.33
0.000
TABLE I: Six-task simulator comparison: 60 trials per method, ten per task. Metrics and scoring subsets are defined in the text.
Configuration
SR (%) ↑
Wait (%) ↓
Calls ↓
SPL ↑
Full A2 (Table I , reused)
73.33
79.40
12.42
0.591
Jev trigger timing
A2 timer-gated Jev (5 s)
65.00
81.83
18.53
0.557
Jev consumption
A2 advisory-only (not consumed)
71.67
79.07
13.25
0.525
Evidence richness
TABLE II: Full A2 and five ablations (60 trials each).
Fig. 3: A2 failure analysis for the same 60-trial cohort as Table I (44 successes, 16 failures). (a) Mutually exclusive recorded failure outcomes, their task composition, and their percentages among all failures. (b) Final distance to the target bounding box and heading error; shading marks the region satisfying both SR thresholds ( ≤1.6 m, ≤60∘ ). Three T3 completion claims cluster at 1.828–1.830 m; one T2 trial fails heading only.
Fig. 5: Successful R5 goal revision (trial 5). Top/bottom rows show scene/robot views at approximately 0, 30, 50, 80, 100, and 120 s. Labels identify the robot and chair/suitcase. The goal changes from chair to suitcase at 22.7 s; panel (c) shows the revised goal, not its issuance. The robot then aligns, approaches, and reaches the suitcase. Timestamped JPEGs use fixed camera crops; times denote frame reception.
Vision-Language Navigation (VLN) requires embodied agents to generate actions based on instructions and observations. General-purpose multimodal agents offer a promising basis for this task, but selecting plausible local actions does not ensure that execution remains consistent with the intended route, particularly in long-horizon tasks. Moreover, the accumulated interaction history increases the input required for subsequent decisions, resulting in a significant inference overhead. To this end, we introduce \method, an Agentic VLN framework that includes a Goal Agent that sets adaptive goals for local actions, a Verify Agent that dynamically verifies whether a goal has been completed, a Memory Agent for multimodal context compression, and a Visuomotor Agent to execute adaptive goals. Specifically, the Goal Agent formulates adaptive goals based on the instruction, current observation, and execution history. Then the Visuomotor Agent executes navigation actions to achieve each goal, while the Verify Agent uses a goal-specific verification question to dynamically assess whether the observed outcomes satisfy the intended completion condition. Verified goal completion then marks a boundary for the Memory Agent to compress the corresponding multimodal interaction history while preserving information needed for subsequent navigation. We evaluate navigation on R2R-CE and RxR-CE, examine framework variants across three model backbones, and study context evolution during execution. For Real-World evaluation, \method achieves 83.3% success and 1.51,m navigation error across eight challenging routes evaluated three times each.
Haoxiang Shi, Zaijing Li, Muhe Ding +3
Harbin Institute of Technology (Shenzhen) · Pengcheng Laboratory
Long-horizon navigation requires an agent to revise its intermediate objectives as evidence accumulates. Full visual histories are costly to process, while compact summaries may omit details needed to reconsider earlier decisions. We introduce NavProbe, a hierarchical zero-shot navigation agent that couples a dynamic subgoal agenda with active evidence retrieval. A compact index links summaries of visited places, transitions, and landmarks to their visual and geometric records. When the current context is insufficient, a task executive retrieves targeted evidence to generate, revise, or resolve subgoals. Reusable conclusions are used to update the index, and a skill policy converts the revised task state into parameterized navigation actions. NavProbe achieves 71.7% SR and 55.8% SPL on R2R-CE and 55.3% SR and 38.6% SPL on RxR-CE, outperforming strong zero-shot baselines. It also achieves 79.3% SR on HM3D-v2 ObjectNav, with qualitative real-robot demonstrations illustrating physical deployment. Code is available at https://github.com/liujy25/NavProbe.
Jingyang Liu, Sujia Yao, Jiayuan Gu +1
School of Information Science and Technology, ShanghaiTech University, Shanghai 201210, China
Agentic navigation systems require a base navigation model whose observation strategy can be externally reconfigured at inference time, because instruction following, object search, target tracking, and autonomous driving share the same perception-planning backbone yet demand fundamentally different strategies for consuming the visual stream. We present Qwen-RobotNav, a scalable navigation model built on Qwen-RobotNav that addresses it through a parameterised interface with two complementary dimensions: multiple task modes that select the navigation behaviour, and controllable observation parameters (e.g., token budget, per-camera weights) that govern how visual history is encoded. With training-time randomization over all parameters, Qwen-RobotNav is robust to any inference-time configuration requiring zero architectural modification to the Qwen-RobotNav backbone. We train Qwen-RobotNav on 15.6M samples; co-training with vision-language data prevents the collapse into reactive action-sequence mappers observed in trajectory-only training. The parameterised interface also makes Qwen-RobotNav a natural building block for agentic systems: for long-horizon scenarios, an upper-level planner decomposes goals into sub-tasks and dynamically switches Qwen-RobotNav's task mode and context strategy mid-episode, composing complex behaviours from repeated calls to the same model. Extensive experiments show that Qwen-RobotNav sets new state-of-the-art results across major navigation benchmarks. The model exhibits favourable scaling from 2B to 8B parameters, with joint multi-task training developing a shared spatial-planning substrate that transfers across task families, and demonstrates strong zero-shot generalisation to real-world robots across diverse environments.