CATNAV: Cached Vision-Language Traversability for Efficient Zero-Shot Robot Navigation

Aditya Potnis; Francisco Affonso; Shreya Gummadi; Naveen Kumar Uppalapati; Girish Chowdhary

cs.ROMar 24, 2026

CATNAV: Cached Vision-Language Traversability for Efficient Zero-Shot Robot Navigation

Authors: Aditya Potnis, Francisco Affonso, Shreya Gummadi, Naveen Kumar Uppalapati, Girish Chowdhary

Organizations: Field Robotics Engineering and Science Hub (FRESH), Illinois Autonomous Farm, University of Illinois at Urbana-Champaign (UIUC), IL

Abstract

Navigating unstructured environments requires assessing traversal risk relative to a robot's physical capabilities, a challenge that varies across embodiments. We present CATNAV, a cost-aware traversability navigation framework that leverages multimodal LLMs for zero-shot, embodiment-aware costmap generation without task-specific training. We introduce a visuosemantic caching mechanism that detects scene novelty and reuses prior risk assessments for semantically similar frames, reducing online VLM queries by 85.7%. Furthermore, we introduce a VLM-based trajectory selection module that evaluates proposals through visual reasoning to choose the safest path given behavioral constraints. We evaluate CATNAV on a quadruped robot across indoor and outdoor unstructured environments, comparing against state-of-the-art vision-language-action baselines. Across five navigation tasks, CATNAV achieves 10 percentage point higher average goal-reaching rate and 33% fewer behavioral constraint violations.

Figures & tables

Method	Traversability		Planning		VLM Efficiency
Method	Zero-Shot Travers.	Dynamic Obj. Segmentation	Motion Prim. Aware Plan.	Multi-Embod. Support	Multi-Modal Costmap	Novelty-Driven VLM Polling	Low-Compute VLM Use
WayFASTER / W-RIZZ / CHUNGUS [ 1 , 2 , 3 ]	–	✓	–	–	–	$\sim$	✓
BEHAV [ 11 ]	$\sim$	–	–	–	✓	–	✓
ConVOI [ 12 ]	$\sim$	–	–	–	–	✓	–
VLM-Social-Nav / Hey Robot [ 13 , 14 ]	$\sim$	–	–	–	–	–	–
VAMOS / OmniVLA [ 15 , 16 ]	$\sim$	–	–	✓	$\sim$	–	–

TABLE I: Comparison of navigation methods against state-of-the-art. ✓ = fully supported,

\sim

= partial/limited, – = not addressed.

Fig. 1: We present CATNAV, a zero-shot cost and embodiment aware traversability navigation method that is able to navigate in unknown environments without fine-tuning.

Fig. 2: Overview of the CATNAV pipeline framework for robot navigation. The system integrates real-time scene perception with risk inference, utilizing a Novelty Check module to determine if a scene requires a new VLM-based costmap prediction or can be processed via cached risks. Following Costmap Construction and Multi-Proposal TRRT Planning, the Trajectory Reasoning module employs a Large Language Model (LLM) to evaluate the proposed paths against specific robot modalities and required behaviors, ultimately selecting the single optimal trajectory for execution.

Panel of Figure 4 — Fig. 3: Scenarios samples: (a) outdoor footpath navigation (Tasks 1–2), (b) outdoor navigation with obstacles (Task 4), (c) dynamic human crossing (Task 3), (d) indoor paper avoidance (Task 5).

	Test 1	Test 2	Test 3
Neighbors ( $k$ )	1	5	5
Novelty Threshold ( $\gamma$ )	0.55	0.1	0.55
$\downarrow$ Scene Queries	16	125	9
$\downarrow$ Path Selector Calls	156	295	157
$\downarrow$ Total VLM Queries	172	420	166
$\uparrow$ Cache Rate (C/s)	0.044	0.104	0.194

TABLE II: VLM Query Rate Comparison Across Configurations.

Fig. 4: Distribution of VLM query frequencies across different caching configurations. The histograms represent the raw query counts, overlaid with Gaussian approximations scaled to the total number of samples and bin width for each test. A reference fixed frequency is plotted based on the fixed polling baseline.

Fig. 5: Qualitative results of the costmap segmentation using CATNAV’s visuosemantic cache. The cost table is aggregated from the $k$ -nearest neighbors of the current scene’s CLIP embedding. If no neighbors are found within the novelty threshold $\gamma$ , the scene is classified as ”Novel,” queried via the LLM, and appended to the vector store.

Method	Goal Reaching(%) $\uparrow$	Dist. to Goal (m) $\downarrow$	Collisions (%) $\downarrow$	Behavior Breaks(%) $\downarrow$
Task 1: Navigate to cone, stay on right side of path (Static)
OmniVLA	60	2.7	20	40
CATNAV (Ours)	70	2.8	10	30
Task 2: Navigate to cone, stay in center of path (Static)
OmniVLA	50	2.3	–	20
CATNAV (Ours)	70	2.25	–	20

TABLE III: Navigation performance of CATNAV vs. OmniVLA across five tasks (10 trials per task)

Panel of Figure 9 — Fig. 6: Trajectory visualizations of CATNAV across all tasks. Colored lines denote candidate and selected trajectories overlaid on the robot’s RGB view., with — representing the baseline and — representing CATNAV .

Robot Traversability Estimation

See all 21 papers

RECAST: Recasting Vision-Language Semantics into an Actionable Cost Map for Robot Navigation
Sep 26, 2026Incheol Cho, Jintae Park, Jinkyu Kim +3Robot Navigation Visual Robot Navigation
Korea University · Hanwha Aerospace · KAIST AI +1
C $^2$ Nav: Compare Before You Commit for Zero-Shot Vision-and-Language Navigation
Sep 14, 2026Runtian Zheng, Congpeng Zhang, Ying LiuVision-Language Navigation VLM Reasoning
North China University of Technology, Beijing, China
One MLLM, One Call: Efficient Zero-Shot Vision-and-Language Navigation via Spatial-Aware Waypoints
Sep 6, 2026Shiqi Pan, Qi Zheng, Hanqin Sun +3Vision-Language Navigation VLMs for Robotics
College of Electronics and Information Engineering, Shenzhen University, Shenzhen 518060, China · School of Computer Science, The University of Sydney, Darlington 2008, Australia · College of Computer Science and Software Engineering, Shenzhen University, Shenzhen 518060, China

Sep 26, 2026J/K move · Enter open · S save

RECAST: Recasting Vision-Language Semantics into an Actionable Cost Map for Robot Navigation

Incheol Cho, Jintae Park, Jinkyu Kim, Jungbeom Lee+2

Korea University · Hanwha Aerospace · KAIST AI+1

Robot Navigation Visual Robot Navigation Safe Robot Navigation VLMs for Robotics

Open paper AskPDF arXiv

Safe and robust robot navigation across diverse environments requires a high-level understanding of complex scenes and the ability to carry it into stable motion. Recent works tackle this with learning-based models trained at scale and with approaches built on vision-language models (VLMs). However, learning-based models break down outside their training distribution, while VLM-based approaches bring that understanding but rarely ground it in the scene or align the action with it. To address these limitations, we present RECAST, a robot navigation framework that combines the reasoning of a VLM with the spatial grounding of vision foundation models to build an Actionable Cost map. Given the robot's front view and the user's instruction, we first decompose the scene with the VLM, judging which surfaces are traversable, which objects pose a risk, which heading to prefer, and which gaps are passable. Vision foundation models then ground these surfaces and objects in the image, and all four judgments are spatially recast into one compact cost map. This map both conditions the trajectory decoders and scores their proposals to select the one to execute. As the VLM's answers trail the live scene, both steps draw on cost maps from two points in time: the pivot frame the VLM judged, which carries all four judgments, and the current frame, whose terrain and collision costs are rebuilt from the current image. RECAST improves success over the strongest prior method by 13.3 points in simulation and 31.4 points on a real quadruped, and reduces the collision rate relative to it by 9.6 and 14.3 points, reaching the lowest collision rate among all methods. The project page is available at https://recast-nav.github.io/

Abstract

Robot Traversability Estimation

Explore similar work

RECAST: Recasting Vision-Language Semantics into an Actionable Cost Map for Robot Navigation

C2^22Nav: Compare Before You Commit for Zero-Shot Vision-and-Language Navigation

One MLLM, One Call: Efficient Zero-Shot Vision-and-Language Navigation via Spatial-Aware Waypoints

C $^2$ Nav: Compare Before You Commit for Zero-Shot Vision-and-Language Navigation