RISMO Steers Drone-Ground Robot Swarms Through Unknown Terrain in Simulation

A team of robotics researchers has unveiled a control framework called RISMO, short for Riemannian Information-Geometric Swarm Motion Orchestration, designed to coordinate mixed teams of aerial and ground robots exploring unknown, GPS-denied environments. The work, published as an open-access original article in the journal Autonomous Intelligent Systems on 10 October 2026, is entirely computational: the framework was evaluated in Matlab simulations of cave-like and confined spaces, not on physical hardware. Even so, the reported results address a persistent headache in multi-robot navigation, namely how to keep a swarm moving smoothly and safely when conventional controllers stall or jitter in tight, non-convex terrain.

The problem the authors target stems from two widely used families of local navigation methods. Artificial potential field (APF) controllers are computationally lightweight and reactive, but in cluttered, non-convex environments they frequently fall into local-minimum deadlocks, where attractive forces toward the goal and repulsive forces from obstacles cancel out and the robots simply stop. Control barrier function quadratic programs (CBF-QPs), by contrast, offer formal safety guarantees for collision avoidance, but the authors report that in their own previous numerical evaluations, repeated activation of hard CBF constraints in narrow corridors produced high-frequency control variation and conservative stagnation, particularly when safety boundaries conflicted with goal-directed motion.

RISMO’s central idea is to fuse perception quality, obstacle proximity, and motion objectives into a single Riemannian metric, a mathematical object that assigns a cost to every possible direction of motion at each point in space. A large metric eigenvalue makes movement along the corresponding direction locally expensive. The framework builds this metric from three ingredients: a baseline inertial term, an obstacle-induced term that inflates the cost of moving near obstacle boundaries, and a perception term derived from the Fisher Information Matrix (FIM), a standard information-theoretic quantity that measures how much a measurement reduces uncertainty. Stronger combined visibility between a ground vehicle and an aerial teammate translates into a lower information-related metric cost, so the resulting geodesic motion naturally favors configurations that preserve cooperative observability.

This perception coupling matters because of the role asymmetry in air-ground teams. UAVs offer agile three-dimensional motion, while UGVs are constrained by ground contact, planar or nonholonomic motion, and narrow-passage clearance. Ground vehicles must simultaneously avoid collisions and maintain high-quality line-of-sight tracking of their aerial teammates, which the relative-localization scheme depends on. In RISMO, a dynamically filtered reference point called the Virtual Agent serves as a shared geometric anchor for the heterogeneous team, and a hierarchical chassis-gimbal coordination mechanism lets the UGV gimbal continuously track the desired line-of-sight bearing while chassis steering is triggered only when gimbal deflection exceeds a physical threshold, preserving observation geometry during maneuvers.

To handle formation changes in narrow passages, the authors model the swarm’s topology as a Semi-Markov Jump System (S-MJS), a stochastic-hybrid framework in which transitions between modes depend on elapsed sojourn time. The system moves among a rigid formation mode, a deformable traversal mode in which the two UGVs reorganize from a lateral pair into a longitudinal queue behind the Virtual Agent, and an escape-assistance mode. Under a positive per-agent inter-event bound and an Average Dwell Time condition, the authors prove that the modeled switching sequence has bounded transition frequency and cannot exhibit Zeno accumulation, the pathological phenomenon in which infinitely many switches pile up in finite time. They further establish practical exponential stability of the switched system, converging to a compact uniformly ultimately bounded set, alongside intra-mode asymptotic stability in obstacle-free space.

A third ingredient tackles symmetric deadlocks directly: a goal-aligned tangential flux that activates near obstacle boundaries and pushes agents along the obstacle tangent rather than into it. The authors prove that at a candidate stagnant state where attractive, inter-agent, and obstacle forces balance to zero, activating this non-conservative tangential component makes the commanded acceleration nonzero, removing the stagnant point from the closed-loop equilibrium set. A chirality operator fixes a local reference orientation, and a stagnation-triggered perturbation supplies tangential input in exactly degenerate cases where the goal-aligned selector would otherwise vanish. A hybrid safety projection additionally removes the inward normal component of commanded velocities near obstacles under finite-step integration.

The headline numerical comparison took place on a procedurally generated 60 by 60 meter cave map with a 0.2 meter grid resolution, nominal corridor radii of 4.0 to 4.8 meters, and 18 circular pillars of radii between 0.3 and 0.8 meters. Three UAVs and two UGVs, all starting at rest, followed a shared upstream target sequence of 71 waypoints under identical actuation limits. RISMO was run over 10 independent S-MJS realizations: nine completed all waypoints within the 750-second horizon, a 90 percent completion rate, while one reached 66 waypoints. Time to 95 percent exploration coverage averaged 563.95 plus or minus 19.88 seconds, mean visibility quality was 0.644 plus or minus 0.030, mean UAV-formation centroid acceleration was 0.0600 plus or minus 0.0029 meters per second squared, and maximum UAV-UGV separation was 15.53 plus or minus 2.83 meters. Complete task-completion time, computed only from the nine successful trials, was 719.42 plus or minus 21.54 seconds.

The two baselines fared worse under the same conditions. The APF controller remained trapped at the first major non-convex obstacle because of conservative force cancellation, while the CBF-QP maintained positive obstacle clearance but did not complete the maneuver within the 750-second horizon and showed larger centroid-acceleration fluctuations. The authors are careful to frame this as a contrast between complementary controller-level paradigms under a common upstream target input, not a ranking of complete navigation architectures; graph-based, learning-based, deadlock-resolution, and MPC methods operate at broader levels and were discussed only at the architecture level. An event audit across the 10 trials recorded zero violations of the 0.02-second minimum inter-event interval for mode transitions, consistent with the Zeno-exclusion analysis.

An ablation study in a compound topological trap, a flat symmetric wall followed by a narrow cylindrical bottleneck with an opening smaller than the rigid formation footprint, isolated each mechanism’s contribution. Removing the tangential flux left the swarm stuck at the flat wall in a projection local minimum, failing to complete the task. Removing the S-MJS deformable mode let the swarm reach the bottleneck but not maintain nominal rigid formation spacing, with absolute inter-UGV spacing deviation exceeding the 0.8-meter rigid-mode tolerance. The full framework entered the deformable mode, reorganizing the UGVs into a sequential queue with a maximum absolute spacing deviation of 14.33 meters while keeping measured obstacle clearance above 0.50 meters, and completed the bottleneck traversal in 118.5 seconds. A three-dimensional evaluation in a 60 by 25 by 8 meter workspace with a V-shaped trap, a converging narrow passage, and suspended obstacles further showed coordinated escape, sequential traversal, and altitude-aware avoidance.

Computation looks feasible for now: on an AMD Ryzen 9 9955HX3D 16-core processor running Matlab R2024b Update 7, the FIM-enabled planar implementation required 2.90 plus or minus 0.02 milliseconds per end-to-end control step across three complete runs, comfortably below the 20-millisecond period of a tested 50-hertz control loop, with six UAV-UGV FIM links evaluated per update. The authors are candid that physical deployment demands more than this study provides, including point-cloud segmentation and data association for relative UAV localization, persistent multi-UAV identity association, and a stable upstream exploration module. They also flag the real-time construction of the perception-dependent Riemannian field from LiDAR observations as the main computational challenge, pointing to lightweight raycasting-based motion policies and learned surrogate models as practical routes forward. Until validation on physical heterogeneous UAV-UGV platforms is reported, RISMO remains a promising, rigorously analyzed simulation framework rather than a field-proven system.

Subject of Research: A Riemannian information-geometric control framework for coordinating heterogeneous UAV-UGV swarms in unknown, GPS-denied environments, evaluated in simulation.

Article Title: RISMO: Riemannian Information-Geometric Swarm Motion Orchestration for Heterogeneous UAV–UGV Cooperative Exploration in Unknown Environments

Article References: Guo, F., Shen, R., Zhang, J., Wang, Y., Sun, C., Wang, Z., & Li, H. (2026). RISMO: Riemannian Information-Geometric Swarm Motion Orchestration for Heterogeneous UAV–UGV Cooperative Exploration in Unknown Environments. Autonomous Intelligent Systems, 6(1), Article 22. https://doi.org/10.1007/s43684-026-00143-2

Image Credits: AI Generated

DOI: 10.1007/s43684-026-00143-2

Keywords: RISMO, UAV-UGV swarms, Riemannian motion policies, Fisher Information Matrix, control barrier functions, artificial potential fields, Semi-Markov Jump Systems, Zeno behavior, obstacle avoidance, cooperative exploration, GPS-denied navigation, simulation study