Source-linked AI summary
FARE: Fast-Slow Agentic Robotic Exploration
Shuhao Liao, Xuxin Lv, Jeric Lew, Shizhe Zhang, Jingsong Liang, Peizhuo Li, Yuhong Cao, Wenjun Wu, Guillaume Sartoretti
TL;DR
Autonomous exploration must acquire geometric information efficiently despite planners’ limited use of long-term structure and fixed adaptation rules. FARE combines LLM-based global reasoning over a pruned belief graph with an RL policy for fast local execution, achieving improved efficiency in simulation and validating full-building exploration on hardware.
Problem
Existing exploration planners struggle to exploit long-term structural information and adapt their strategies to evolving or different environments.
Method
FARE uses a hierarchical fast-slow architecture in which an LLM generates global guidance from environment descriptions and a pruned graph, while an RL policy executes locally using graph, utility, and waypoint information.
Results
FARE improves exploration efficiency over baselines in challenging simulated environments and successfully explores an entire 200m × 130m indoor building without manual intervention.
Takeaways & Limitations
Separating environment-conditioned global reasoning from fast local decision-making yields coherent exploration that reduces redundant backtracking while retaining local flexibility.
Takeaways & Limitations
Future work targets multi-robot coordination, vision-based semantic perception, richer representations, and three-dimensional action spaces to improve generality.
Abstract
from arXiv · showhide
This work advances autonomous robot exploration by integrating agent-level semantic reasoning with fast local control. We introduce FARE, a hierarchical autonomous exploration framework that integrates a large language model (LLM) for global reasoning with a reinforcement learning (RL) policy for local decision making. FARE follows a fast-slow thinking paradigm. The slow-thinking LLM module interprets a concise textual description of the unknown environment and synthesizes an agent-level exploration strategy, which is then grounded into a sequence of global waypoints through a topological graph. To further improve reasoning efficiency, this module employs a modularity-based pruning mechanism that reduces redundant graph structures. The fast-thinking RL module executes exploration by reacting to local observations while being guided by the LLM-generated global waypoints. The RL policy is additionally shaped by a reward term that encourages adherence to the global waypoints, enabling coherent and robust closed-loop behavior. This architecture decouples semantic reasoning from geometric decision, allowing each module to operate in its appropriate temporal and spatial scale. In challenging simulated environments, our results show that FARE achieves substantial improvements in exploration efficiency over state-of-the-art baselines. We further deploy FARE on hardware and validate it in complex, large scale $200m\times130m$ building environment.
I. INTRODUCTION
Autonomous exploration seeks efficient information acquisition in unknown environments, but existing planners struggle to use long-term structure and adapt beyond fixed heuristics. FARE addresses this with hierarchical fast-slow reasoning that separates global strategy from local decisions.
- Autonomous exploration enables mobile robots to acquire geometric information while incrementally building environmental maps.Sensors such as 3D LiDARs or cameras support representations including occupancy grids and voxel maps.
- Conventional methods use geometric belief maps and hierarchical coarse-to-fine planning to balance discovering new regions with refining partially explored areas.These approaches compute low-resolution global paths and high-resolution local paths, but their decomposition and parameters remain predefined.
- Existing planners struggle to exploit long-term structural information and adapt exploration strategies to different environment characteristics.They often rely on fixed heuristics, rigid rules, or constant hyperparameters governing path length, information gain, and representation resolution.
- FARE separates global reasoning from local decision making through a hierarchical fast-slow thinking framework.Its slow module reasons globally over environmental structure, while the fast module performs local decisions.
- Learning-based approaches formulate exploration as sequential decision making and estimate long-term returns from observations.Many use convolutional neural networks over fixed-size maps or local occupancy grids, with actions ranging from viewpoint selection to direct navigation commands.
B. LLM-based Graph Reasoning
The paper frames exploration as incremental reasoning over a partially observed occupancy map, where the robot must complete mapping with a collision-free, low-cost trajectory. The formulation distinguishes known, unknown, free, and occupied regions.
- The environment is a bounded, initially unknown two-dimensional space represented by an occupancy grid.The robot incrementally constructs a partial map during exploration.
- The partial map is decomposed into known and unknown regions, with the known region divided into traversable free space and occupied space.The decomposition is expressed as M = M_u ∪ M_k and M_k = M_f ∪ M_o.
- At each decision step, an omnidirectional LiDAR classifies cells within sensing range as free or occupied according to traversability.The sensing range is denoted d_s.
- Autonomous exploration computes a collision-free trajectory that minimizes total traversal cost while completing exploration.In simulation and benchmarks, the ground-truth map supports evaluation, while completion is commonly approximated by no remaining unknown or frontier regions.
IV. METHODOLOGY
FARE represents local robot belief with a collision-free graph built from sensor data and uses it to support viewpoint selection. Local candidate nodes receive frontier-based utility and are restricted to a sliding spatial window.
- 1) Local Belief Graph:: The robot belief is represented by a collision-free graph constructed from onboard sensor data.The trajectory is a sequence of free-space viewpoints, τ = (v_0, v_1, ...).
- 1) Local Belief Graph:: Candidate viewpoints are sampled from current free space, connected to nearby neighbors, and pruned when edges intersect occupied or unknown regions.This yields a collision-free graph G_t = (V_t, E_t).
- 1) Local Belief Graph:: Each viewpoint receives utility equal to the number of observable frontiers within sensor range.A frontier is observable when its connecting segment is collision-free and its distance from the viewpoint does not exceed d_s.
- 1) Local Belief Graph:: A square sliding window centered at the robot position defines the local belief graph used for nearby decision making.The local graph contains candidate nodes and connecting edges within the window.
2) Global Belief Graph with Modularity-Based Pruning:
FARE constructs a compact global belief graph by detecting communities in the collision-free graph and retaining only structurally informative ones. Modularity-based top-k pruning reduces reasoning complexity while preserving dominant topological structure.
- 2) Global Belief Graph with Modularity-Based Pruning:: The global belief graph is built by combining community detection with modularity-based pruning on the collision-free graph.Only selected communities are promoted to high-level nodes for global reasoning.
- 2) Global Belief Graph with Modularity-Based Pruning:: Graph modularity is defined using edge count, adjacency, node degrees, and community membership indicators.The formulation uses m, A_ij, k_i, and δ(·) to quantify community structure.
- 2) Global Belief Graph with Modularity-Based Pruning:: The community-level modularity objective ignores edge weights and directions while reorganizing the objective over detected communities.For each community, P_in counts internal edges and P_tot counts edges connected to it.
- 2) Global Belief Graph with Modularity-Based Pruning:: FARE retains only the top-k communities with the highest modularity contributions.This pruning step is integrated directly into global graph construction.
- 2) Global Belief Graph with Modularity-Based Pruning:: Each retained community becomes a global node, with edges induced by inter-community connectivity.The resulting graph is compact yet informative, reducing reasoning complexity while preserving dominant environmental topology.
B. Slow-Thinking Module
The slow-thinking module converts a concise natural-language environment description into structured characteristics and an environment-conditioned exploration strategy. The strategy is parameterized across spatial, efficiency, safety, and task dimensions.
- Strategy Synthesis: The module translates high-level environment descriptions into strategy-level guidance rather than directly prescribing actions or trajectories.This bridges semantic understanding with long-horizon exploration planning.
- Environment Characterization: The LLM characterizes environments using spatial characteristics, obstacle characteristics, and exploration challenges.The predefined schema includes layout, obstacle properties, and navigation-related difficulties.
- Strategy Synthesis: The exploration strategy is parameterized by spatial, efficiency, safety, and task dimensions.These axes encode coverage and traversal order, energy and time trade-offs, safety behavior, and task objectives.
- Strategy Synthesis: A strategy instance can specify boundary-first coverage, perimeter following, balanced depth–breadth, conservative energy use, and conservative obstacle clearance.The example also sets moderate time constraints and backtrack tolerance while avoiding revisits.
2) Graph Reasoning:
Graph reasoning uses pruned global graph structure, a textual strategy prompt, and episode memory to iteratively generate a path toward unexplored areas.
- Graph Reasoning: The LLM ensemble performs iterative reasoning over the pruned global belief graph using a textual strategy prompt and episode memory.Reasoning proceeds at successive depths i.
- Graph Reasoning: The resulting global path τg begins at the current node and terminates at an unexplored node.It is represented as a sequence containing the current node followed by intermediate nodes and a target node.
C. Fast-Thinking Module
The fast-thinking module represents local exploration with an informative graph containing node positions, utility, and global-path guideposts, then selects neighboring waypoints through graph attention.
- Policy Network: The policy observation integrates the local graph, utility, and global path as ot = (G*, τt).This combines local geometric information with long-horizon guidance.
- Policy Network: The fast-thinking policy selects a neighboring node as the next waypoint and executes motion toward it.The informative graph augments local node positions with utility and a binary guidepost indicating membership in global paths.
- Policy Network: Attention weights are normalized with an adjacency-based mask so aggregation follows current graph-edge constraints.Each output node feature aggregates weighted values across neighboring nodes.
- Policy Network: Figure 2 distinguishes the robot position, policy-selected waypoint, and globally guided next waypoint at timestep t.The illustration shows how local waypoint selection relates to global guidance.
2) Instruction Following:
Instruction following trains the fast-thinking policy to remain aligned with global waypoint guidance through a smooth deviation-based reward. This is intended to reduce detours while preserving reinforcement-learning compatibility.
- Instruction Following: The fast-thinking policy selects a local waypoint wt while following a global path whose next waypoint is w*t.The instruction-following objective encourages adherence to long-horizon guidance.
- Instruction Following: The normalized deviation between the selected and globally guided waypoints provides a smooth surrogate for path-length deviation.The surrogate preserves a monotonic relationship with deviation from global guidance.
- Instruction Following: The exponential instruction-following penalty mildly penalizes small deviations and increasingly suppresses large deviations.This produces smooth and stable gradients for policy optimization.
- Evaluation: Figure 3 compares FARE with conventional baselines across 10 runs per method in each environment.The figure is a comparative evaluation of exploration performance.
- Evaluation: Table I reports comparison results across three environments.The supplied caption identifies the table’s scope but does not specify individual metrics or values.
- Evaluation: Figure 4 shows FARE exploration trajectories in indoor and outdoor Gazebo simulations using color to encode movement over time.The visualization emphasizes trajectory progression across the simulations.
- Instruction Following: Minimizing cumulative instruction-following penalties implicitly encourages shorter executed trajectories while remaining compatible with efficient reinforcement learning.The paper connects waypoint alignment to reduced detours and trajectory length.
V. EXPERIMENTS
FARE is evaluated for effectiveness and robustness across diverse simulated environments and a large-scale real mobile-robot deployment.
- Experiments compare FARE across indoor, forest, and warehouse Gazebo environments, followed by validation on a real mobile robot in a large-scale campus environment.The evaluations use the same trained model and identical system configurations unless otherwise specified.
A. Comparison Analysis
FARE matches baselines in compact indoor spaces but gains increasing efficiency in structurally distinctive forest and warehouse environments through global structural guidance and adaptive coordination with local execution.
- FARE performs on par with other baselines in compact indoor environments lacking distinctive global-level structure.
- In forest environments, FARE reduces travel distance and makespan, with its advantage widening in warehouse environments.
- FARE achieves the shortest paths and fastest completion times among all methods in warehouse environments.
- FARE completes peripheral and corner regions early, whereas baselines often defer them and incur additional backtracking.The trajectory comparison indicates that FARE incorporates global structural cues rather than relying solely on local frontier or utility signals.
- FARE’s gains are attributed to environment-adaptive planning and coordination between long-horizon global guidance and flexible local execution.The slow-thinking module provides global direction while the fast-thinking module exploits nearby informative regions.
B. Hardware Validation
FARE transfers to hardware by using onboard LLM inference to explore a large indoor teaching building without manual intervention, while future work targets broader deployment settings.
- FARE runs on an Agilex Scout-mini robot with onboard Qwen3-14B inference, Ouster OS0-32 LiDAR perception, and FastLIO2 for odometry and mapping.The deployment uses a maximum speed of 1 m/s, an 8 m sensor range, and 1 Hz local replanning.
- FARE successfully explores an entire 200m × 130m indoor teaching building without manual intervention while maintaining stable runtime performance.The building contains long corridors, rooms, and intersections, and the map and node resolutions remain fixed during deployment.
- The framework combines environment-conditioned global reasoning, pruned-graph guidance, and fast local decision-making to reduce redundant backtracking in real-world exploration.
- Future work extends FARE toward multi-robot coordination, online environment-change detection, richer representations, and three-dimensional action spaces.These directions aim to support hybrid environments with abrupt scene-type transitions and improve generality.