MM-Nav: Multimodal Navigation with Multi-Expert Learning

• MM-Nav is a class of multimodal models that fuse visual, language, and action cues for robust navigation in diverse environments.
• It employs a multi-expert teacher–student framework to learn continuous control, integrating multi-view inputs with dynamic corrective feedback.
• A minimalist variant, MinNav, uses monocular optical flow and uncertainty measures to enable efficient navigation for tiny aerial robots.

MM-Nav refers to a class of multimodal and multi-expert models for embodied visual navigation that integrate vision, language, and action. The term denotes both a specific state-of-the-art method for multiview vision-language-action learning in robotics (Xu et al., 3 Oct 2025), as well as a separate minimalist navigation system for tiny aerial robots based on monocular optical flow and uncertainty (Patil et al., 5 Jun 2026). This article provides a comprehensive account of MM-Nav in both senses, with an emphasis on the prominent multi-view VLA framework for robust navigation via multi-expert learning.

1. Model Definition and Core Architectural Elements

MM-Nav, in the context of robust visual navigation, is an end-to-end Vision-Language-Action (VLA) model that maps 360° egocentric RGB observations and a point-goal prompt to continuous velocity commands $(v_x,v_y,v_{yaw})$ (Xu et al., 3 Oct 2025). The architecture is composed of three primary modules:

• Multi-view Visual Encoder: Utilizes four horizontal fisheye cameras (Front, Right, Back, Left), each providing an eight-frame temporal history. Images are patch-tokenized using a high-capacity visual foundation model (SigLIP), reduced via a learned two-layer MLP projector to produce a token sequence of shape $\mathbb{R}^{192\times C}$.
• Language Reasoning Module: Converts the relative point-goal into a textual prompt, encoded with the 7B-parameter Qwen2 LLM. The prompt and visual tokens are concatenated and jointly processed, yielding an “action token.”
• Action Decoder: A two-layer MLP maps the LLM’s action token to continuous velocity commands.

The model is trained on a composite loss of mean squared error on velocity prediction and cross-entropy loss on open-world VQA questions.

2. Multi-Expert Teacher–Student Training Paradigm

MM-Nav employs a teacher–student framework, where three privileged RL “teacher” experts are trained using depth observations and tailored reward functions to specialize in reaching, squeezing, and dynamic obstacle avoidance. Each teacher is a PPO agent with depth inputs and proprioceptive state, trained in a custom environment that targets its capability (e.g., narrow corridor navigation for squeezing, moving obstacle fields for avoiding).

The VLA student learns in two phases:

• Offline behavior cloning: The student is initialized by behavior cloning on the aggregate expert dataset ($500k$ trajectories).
• Online DAgger-style iteration: The student is deployed in all navigation tasks, with corresponding experts providing corrective actions for aggregation and dynamic rebalancing.

Crucially, data proportions for each expert are dynamically balanced according to a performance-gap metric based on Weighted Travel Time to encourage focused improvement on the weakest skill in each iteration.

3. Synthetic and Real-World Experimental Protocols

Experiments are conducted in simulation (IsaacLab), using three challenge environments and a fourth mixed-capability scene. Training and evaluation protocols are as follows:

• Episode Termination: Success, collision, or timeout.
• Data and Compute: Offline stage uses $500k$ expert steps and $100k$ VQA steps; each online iteration uses $200k$ expert and $40k$ VQA steps on 8 H100 GPUs.
• Observation Space: VLA student receives RGB images; RL teachers use depth.
• Metrics: Success Rate (SR), Collision Rate (CR), and Weighted Travel Time (WTT).

In real-world trials, the model demonstrates zero-shot transfer on a Unitree GO2 quadruped with surround-view fisheye cameras at 7 Hz.

4. Quantitative Benchmarks and Comparative Evaluation

The following summarizes synthetic benchmarks from (Xu et al., 3 Oct 2025):

Method	Reaching (SR/CR/WTT)	Squeezing (SR/CR/WTT)	Avoiding (SR/CR/WTT)	Mixed (SR/CR/WTT)
iPlanner	19/81/93.4	2/94/881.3	15/85/73.4	4/94/736.9
ViPlanner	43/57/42.2	4/96/452.8	22/78/36.4	18/82/215.4
NavDP	69/31/27.3	18/82/115.4	27/73/30.0	23/77/178.6
MM-Nav	80/20/31.0	71/19/42.2	68/32/20.9	47/26/127.5

Ablation indicates combining all expert datasets (mixed-VLA) leads to superior generalization: Reaching/Squeezing/Avoiding single-task models are outperformed by the mixed-expert VLA in their respective domains.

Qualitative trials on hardware validate the approach: MM-Nav successfully executes tasks such as squeezing through narrow boxes, avoiding dynamic or transparent obstacles, and exhibiting zero-shot robustness in real cluttered scenes.

5. Minimalist MM-Nav for Tiny Aerial Robots (MinNav)

A distinct usage of “MM-Nav” appears in (Patil et al., 5 Jun 2026) describing MinNav, a lightweight monocular navigation system for tiny UAVs employing dense optical flow and flow-uncertainty for scene-understanding and collision avoidance. The pipeline includes:

• Flow and uncertainty estimation via a 2.8 M parameter pyramidal CNN.
• Scene-motif classification for gap detection and dynamic obstacle avoidance.
• Active exploratory strategies to escape focus-of-expansion dead zones and maximize free-space discovery.
• Experimental validation on a Corgi210 quadrotor shows real-world success rates of 70%, comparable to depth-based alternatives at lower computational cost.

This approach is orthogonal to VLA-based MM-Nav: it emphasizes minimalist, perceptually-driven navigation without reliance on language or privileged depth.

6. Extensions, Insights, and Limitations

Insights

• Multi-view learning synergistically integrates multiple navigation skills, enabling implicit geometric reasoning from RGB alone and surpassing the capacity of privileged-depth teachers.
• Shared representations in LLM backbones facilitate robust capability fusion and regularization via open-world QA, improving generalization.

Limitations

• MM-Nav VLA inference is currently limited by compute (∼7 Hz), restricting real-time deployment on resource-constrained robots.
• Multi-view coverage is restricted to the horizontal plane; vertical navigation challenges remain untested.
• Extension to cross-embodiment tasks (drones, legged, dynamic following) and to environments with novel physical characteristics (stairs, ramps) is ongoing.

Future Directions

• Residual RL for cross-embodiment adaptation (Xu et al., 3 Oct 2025).
• Domain adaptation for visual appearance/geometry shifts.
• Integration of additional expert skills (e.g., human following, social navigation).
• Further network optimization for embedded deployment.
• For MinNav: acceleration via optimized inference libraries, event-camera integration, and hybridization with stereo/depth where feasible (Patil et al., 5 Jun 2026).

7. Relationship to Related Multimodal Navigation Frameworks

GridMM (Wang et al., 2023) and NavBench (Qiao et al., 1 Jun 2025) situate MM-Nav within a landscape of vision-language navigation and benchmarking approaches. GridMM uses a top-down egocentric grid for fine-grained instruction-relevant memory, whereas MM-Nav fuses multi-view egocentric vision with language for continuous control. NavBench illustrates the integration of MLLMs with robotic actuation, highlighting the challenge of temporal alignment and execution, which remain open problems for MM-Nav and similar systems.

In summary, MM-Nav designates a unified approach to multimodal navigation that leverages multi-expert learning, scalable visual-language representation, and cross-domain generalization, achieving state-of-the-art performance in both simulation and real-world mobile robotics settings (Xu et al., 3 Oct 2025, Patil et al., 5 Jun 2026).