NavMorph: Adaptive Navigation & Planning

Updated 22 May 2026

NavMorph is a suite of adaptive navigation frameworks that integrates latent-dynamics modeling with vision-and-language navigation and MPC-based control.
It employs self-evolving world models and Contextual Evolution Memory to enable real-time prediction, online adaptability, and improved performance metrics.
The system combines hybrid 2D/3D planning with precise MPC strategies to achieve collision-free, efficient navigation in ultra-narrow, complex spaces.

NavMorph refers to multiple advanced frameworks for navigation and planning in autonomous systems, encompassing: (1) a self-evolving world model for Vision-and-Language Navigation in Continuous Environments; (2) a hybrid planning and control suite for morphogenetic UAVs combining bi-modal A* navigation and gait-level MPC; and (3) a constrained MPC-based planning solution for morphing aerial robots in ultra-narrow spaces. While distinct in application and architectural specifics, these systems share an emphasis on online adaptability, latent-dynamics modeling, and closed-loop interaction with perception or language input. Below, the main scientific lines of NavMorph are synthesized, tracing their formalism, implementations, and performance.

1. Motivations and Problem Spaces

NavMorph encompasses frameworks addressing navigation in (a) vision-and-language navigation in continuous 3D environments; (b) hybrid terrestrial/aerial morphogenetic robots; and (c) aerial robotics maneuvering through constrained geometry. The motivation underlying all variants is to endow robotic agents with robust adaptability to dynamic or previously unseen environments, avoid reliance on precomputed maps or static reasoning, and operate across unreliable observations, morphing modalities, or ambiguous instructions.

In VLN-CE, the core challenge is the execution of navigation actions from free-form instructions, generating continuous low-level motion primitives under natural language constraints, continuous dynamics, and partial observability. For morphogenetic UAVs and morphing quadrotors, the challenge lies in real-time, collision-free, and energy-efficient traversal of environments requiring on-the-fly morphological and mode transitions, such as folding/unfolding limbs or alternating between ground and flight (Yao et al., 30 Jun 2025, Mustafa et al., 2024, Modi et al., 15 May 2026).

2. Self-Evolving World Models in VLN-CE

The NavMorph approach to VLN-CE is based on a recurrent state-space model (RSSM) with a self-evolving memory. The architecture consists of two principal modules:

World-Aware Navigator (Inference Network): Infers latent environmental state from RGB-D observations and action histories. At time $t$ , the hidden state $h_t$ is updated with $h_t = f(h_{t-1}, s_{t-1})$ , and a stochastic latent $s_t$ is inferred via

$q_\phi(s_t \mid o_{1:t},a_{1:t-1}) = \mathcal{N}(\mu_\phi(h_t,a_{t-1},x_t), \sigma_\phi(h_t,a_{t-1},x_t) I).$

Foresight Action Planner (Predictive Network): Rolls out imagined latent sequences for planning. The prior and predicted trajectory are given by

$p_\theta(\hat{s}_t \mid h_t, \hat{a}_{t-1}) = \mathcal{N}(\mu_\theta(h_t,\hat{a}_{t-1}), \sigma_\theta(h_t,\hat{a}_{t-1}) I),$

where $\hat{a}_t \sim \pi_\theta(\hat{h}_t, \hat{s}_t)$ with $\hat{h}_t$ updated as above.

Learning is based on maximizing a variational lower bound, regularized with NDTW-based trajectory similarity losses, and includes an imitation learning term to enforce intended behavior:

$\mathcal{L}_W = \sum_{t=1}^T \mathbb{E}[ \ell_{re}_t + \ell_{ac}_t ] + \sum_{j=1}^{T_p} \mathbb{E}[ \ell_{re}_{T+j} + \ell_{ac}_{T+j} ] + \gamma \sum_{t=1}^T \mathbb{E}\left[ D_{KL}( q_\phi(s_t) \| p_\theta(s_t) ) \right].$

The Contextual Evolution Memory (CEM) is a fast, non-gradient-evolving bank of contextual vectors. At each step, the top-K similar vectors are retrieved for contextual blending, then updated in place with blending factors to promote memory adaptation without full parameter retraining. This accelerates online adaptation to novel scene contexts and environment dynamics (Yao et al., 30 Jun 2025).

NavMorph, as introduced for morphogenetic UAVs (MorphoGear), integrates a hierarchical planning-control paradigm:

Hybrid Bi-Modal Path Planner: The environment is discretized into 2D and 3D grids. The planner first attempts collision-free pathfinding in 2D (ground); failing that, it transitions to 3D (flight) with an explicit cost penalizing mode changes and a "landing bonus" to promote return to ground. Formally, the cost accumulates as:

$g(j) = g(i) + \Delta d(i, j) + \delta_{\text{mode}}(i \to j),$

with $h_t$ 0 capturing entry/exit transition costs.

MPC-Based Ground Locomotion Follower: The ground follower solves a finite-horizon QP using a kinematic template of the robot's canter gait. The cost function penalizes tracking error (with $h_t$ 1) and control effort ( $h_t$ 2), subject to actuation and collision constraints. The model predictive control is robust, enabling sub-centimeter root mean squared error (RMSE $h_t$ 3 cm) in simulated ground path-following (Mustafa et al., 2024).

4. Constrained MPC for Morphing Quadrotors in Ultra-Narrow Passages

An alternative NavMorph instantiation targets real-time trajectory planning for shape-adaptive UAVs navigating tight environments, using nonlinear MPC with integrated morphing control (Modi et al., 15 May 2026):

State/Actuation Model: The 14-dimensional state $h_t$ 4 spans positions, velocities, angles, angular rates, and morphing angles/rates; inputs $h_t$ 5 include motor thrusts and morphing actuator torque.
Obstacle-Avoidance Cost Function: A novel exponential potential function without hard activation cutoffs is defined:

$h_t$ 6

where $h_t$ 7 is the minimum distance to obstacle segments (derived from clustered and split 2D LiDAR input); $h_t$ 8 is a narrow-gap cost reduction factor favoring centerline passage within narrow gaps.

Perception and Planning Association: 2D LiDAR scans are processed via DBSCAN clustering and line regression; the resulting geometric primitives are directly used for minimum distance queries in the MPC cost at each prediction step. No explicit occupancy grid or nonlinear geometric constraints are imposed—only proximity measures to line segments appear in the cost (Modi et al., 15 May 2026).
Performance: The framework enables traversal through gaps as narrow as $h_t$ 9– $h_t = f(h_{t-1}, s_{t-1})$ 0 m in both simulation and hardware, maintaining full pass rates and rapid response, with solve times under $h_t = f(h_{t-1}, s_{t-1})$ 1 s per MPC iteration.

5. Empirical Results and Comparative Performance

The three NavMorph instantiations share an emphasis on empirical validation with rigorous metrics:

Application	Key Error Metrics	Adaptivity Features	Noted Limitations
VLN-CE World Model (Yao et al., 30 Jun 2025)	R2R-CE Val Unseen: NE ↓ (6.05→5.75), SR ↑ (43.8→47.9), SPL ↑ (29.4→33.2)	Online latent memory (CEM), foresight rollouts	Rollout horizon $h_t = f(h_{t-1}, s_{t-1})$ 2 degrades SPL/SR
MorphoGear Hybrid Nav (Mustafa et al., 2024)	RMSE 0.91 cm, Max 1.85 cm (MPC on ground); flight/ground switching	Bidirectional mode transitions, hybrid A* + MPC	Static maps, ignores flight stability
Morphing Quadrotor MPC (Modi et al., 15 May 2026)	100% success in $h_t = f(h_{t-1}, s_{t-1})$ 3 m gap, deviation $h_t = f(h_{t-1}, s_{t-1})$ 4 m (with $h_t = f(h_{t-1}, s_{t-1})$ 5 term)	Real-time shape planning, perception-driven MPC	Limited by LiDAR FOV

In VLN-CE, NavMorph outperforms prior baselines in navigation error and success, with ablation confirming the necessity of CEM and multi-term loss. In hybrid UAV navigation, the system yields a $h_t = f(h_{t-1}, s_{t-1})$ 691% improvement in path-tracking over open-loop. For morphing UAVs, the new cost function enables gap-traversal outcomes unattainable with APF-style costs.

6. Systemic Limitations and Prospects

Common limitations across NavMorph implementations include assumptions of static or fully observable environments, computational scaling of grid-based planners or large memory, and restricted actuation models (e.g., fixed yaw, idealized aerial control). Prospective improvements, noted in the source works, include:

Adoption of sampling-based planners (RRT*, PRM) or incremental strategies (D*, LPA*) for larger or dynamic spaces (Mustafa et al., 2024).
Integration of full nonlinear limb/body dynamics, MPC-based flight for aggressive maneuvers (Mustafa et al., 2024).
Fusion of onboard vision with online map evolution for unified mode/policy optimization (Mustafa et al., 2024).
Applicability of the exponential+μ cost function to arbitrary mobile robots, beyond morphing quadrotors (Modi et al., 15 May 2026).
Extension of evolving memory and latent-dynamics models to multi-agent or more complex perception-action loops (Yao et al., 30 Jun 2025).

7. Implementation and Reproducibility

Codes and technical artifacts for the respective NavMorph variants are made available by the respective research groups. VLN-CE world-model code is hosted at https://github.com/Feliciaxyao/NavMorph (Yao et al., 30 Jun 2025), while the morphing quadrotor MPC framework is available at https://github.com/harshjmodi1996/morphocopter_mpc (Modi et al., 15 May 2026). Deployment pipelines routinely rely on CasADi/acados, Unity or Gazebo simulators, ROS nodes, and hardware abstraction interfaces, with stated hyperparameters and learning schedules specified for ready reproduction.

NavMorph thus represents a family of principled, generalizable, and perception-/dynamics-integrated frameworks for navigation in complex, unconstrained, or morphologically adaptive environments, with broad applicability across language-guided agents, hybrid UAVs, and shape-shifting robots. The core unifying concept is the explicit modeling of latent environmental structure and agent-environment interaction in a framework supporting both foresightful planning and continual online adaptation.