💻 computer science

FARE: Fast-Slow Agentic Robotic Exploration

FARE is a hierarchical autonomous exploration framework that integrates a large language model for global semantic reasoning and a reinforcement learning policy for fast local control, enabling efficient and robust navigation in complex, large-scale environments.

Original authors: Shuhao Liao, Xuxin Lv, Jeric Lew, Shizhe Zhang, Jingsong Liang, Peizhuo Li, Yuhong Cao, Wenjun Wu, Guillaume Sartoretti

Published 2026-01-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shuhao Liao, Xuxin Lv, Jeric Lew, Shizhe Zhang, Jingsong Liang, Peizhuo Li, Yuhong Cao, Wenjun Wu, Guillaume Sartoretti

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are sending a robot into a giant, completely unknown building to map it out. The challenge is: How do you get the robot to explore efficiently without getting lost, wasting time, or running in circles?

The paper introduces FARE (Fast-Slow Agentic Robotic Exploration), a new way to program robots that mimics how humans think: by combining slow, big-picture thinking with fast, quick reactions.

Here is how it works, broken down into simple concepts:

1. The Two Brains: The Architect and the Driver

FARE splits the robot's "brain" into two distinct parts that work together, like a Team of an Architect and a Driver.

The "Slow" Brain (The Architect): This part uses a Large Language Model (AI that understands text). It doesn't look at every single brick; instead, it looks at a simplified "map" of the building and reads a short description (e.g., "This is a warehouse with narrow aisles and tall shelves").
- What it does: It acts like a master planner. It figures out the big strategy: "Okay, we should go to the back corner first because it's a dead end," or "Let's stick to the perimeter." It creates a list of global checkpoints (waypoints) to visit.
- The Trick: To keep this fast, the AI doesn't look at the whole messy map. It uses a "pruning" technique, like a gardener trimming a bush, to cut away unnecessary details and only keep the most important structural parts of the map.
The "Fast" Brain (The Driver): This part uses Reinforcement Learning (a type of AI that learns by trial and error). It is the robot's eyes and feet.
- What it does: It reacts instantly to what it sees right in front of it. If there is a chair or a wall, it steers around it immediately.
- The Connection: Even though the Driver is fast, it isn't wandering aimlessly. It is constantly checking the Architect's list of global checkpoints. It tries to drive toward the next checkpoint while dodging local obstacles.

2. The "Fast-Slow" Analogy

Think of driving a car on a long road trip:

Slow Thinking: Before you leave, you look at a map app and plan your route. You decide, "I need to take Highway 95 to get to the next city." This is the global strategy.
Fast Thinking: While driving, you are constantly reacting to the road. You swerve to avoid a pothole, slow down for a deer, or change lanes to avoid a slow truck. This is the local reaction.

Older robots often only had the "Fast" part. They would drive until they hit a wall, then turn, then drive again. They didn't have a long-term plan, so they often ended up backtracking (driving back over the same ground) or getting stuck in loops. FARE gives the robot the "Slow" planner so it knows where it needs to go in the long run, while the "Fast" driver handles the how.

3. How They Talk to Each Other

The paper describes a special "reward system" to make sure the Fast Driver listens to the Slow Architect.

If the Driver starts to drift too far away from the Architect's planned path, it gets a "penalty" (a negative signal).
If it stays on track, it gets a "reward."
This teaches the robot to balance being flexible enough to dodge obstacles but disciplined enough to follow the big plan.

4. The Results: Does It Work?

The researchers tested FARE in three different "worlds":

Indoor Office: A building with long hallways and rooms.
Forest: An outdoor area with trees and uneven ground.
Warehouse: A place packed with tall stacks of boxes and narrow aisles.

The Findings:

In simple, open spaces, FARE performed just as well as the best existing robots.
In complex, cluttered places (like the Forest and Warehouse), FARE was significantly better. It traveled shorter distances and finished the job faster.
Why? Because it didn't waste time going back and forth. It knew early on which areas were "dead ends" and tackled them first, rather than leaving them for last.

5. Real-World Test

The team didn't just test this in a computer simulation. They put the system on a real, wheeled robot (an Agilex Scout-mini) and drove it through a massive 200m x 130m university teaching building.

The robot successfully mapped the entire building on its own.
It handled the transition from the computer simulation to the real world without needing to be re-tuned, proving the system is robust.

Summary

FARE is a robot explorer that stops guessing and starts planning. By giving the robot a "slow thinker" to map out the big picture and a "fast thinker" to handle the immediate driving, it explores new environments much more efficiently, avoiding the wasted steps that plague older robots.

Technical Summary: FARE (Fast-Slow Agentic Robotic Exploration)

Problem Statement

Autonomous robot exploration in unknown environments aims to efficiently acquire geometric information to build a complete map. While modern Simultaneous Localization and Mapping (SLAM) and LiDAR odometry have solved localization and mapping reliability, the exploration planning problem remains challenging in complex, real-world settings.

Existing approaches face two primary limitations:

Conventional Planners: These methods often rely on fixed heuristics or rigid decision rules (e.g., balancing path length vs. information gain) that do not adapt to varying environmental characteristics. They struggle to leverage long-term structural information embedded in evolving maps, leading to suboptimal behaviors such as excessive backtracking or unnecessary refinement of already explored areas.
Learning-Based Methods: Reinforcement Learning (RL) approaches often suffer from sparse reward signals, as the objective (minimizing total time/distance) is only observable at the end of an episode. While dense rewards based on immediate information gain improve training stability, they bias the policy toward short-term gains, making it difficult to learn long-horizon behaviors like strategic backtracking or exploiting distant structural cues.

Current methods generally fail to effectively utilize long-term information or adapt their exploration strategies as environmental conditions change dynamically.

Methodology

The paper proposes FARE (Fast-Slow Agentic Robotic Exploration), a hierarchical framework that decouples semantic reasoning from geometric decision-making by integrating a "slow-thinking" Large Language Model (LLM) module with a "fast-thinking" Reinforcement Learning (RL) module.

1. Hierarchical Robot Belief Graph

FARE constructs a dual-layer graph representation of the environment:

Local Belief Graph ( $G_{local}$ ): A dense, collision-free graph derived from onboard sensor data (LiDAR) within a sliding window around the robot. Nodes represent candidate viewpoints, and edges connect neighbors, serving as input for the fast-thinking module.
Global Belief Graph ( $G_{global}$ ): A sparse graph constructed by applying modularity-based community detection to the full map. A pruning mechanism retains only the top- $k$ communities with the highest modularity contributions. These communities are abstracted as high-level nodes, creating a compact representation of the environment's topological structure for global reasoning.

2. Slow-Thinking Module (Global Reasoning)

This module operates at a high temporal and spatial scale to generate strategic guidance:

Environment Characterization: The LLM receives a concise natural language description of the environment (e.g., "confined corridors," "irregular obstacles"). It analyzes these inputs to extract spatial, obstacle, and challenge characteristics.
Strategy Generation: Based on the characterization, the LLM instantiates a structured exploration strategy parameterized along four axes: spatial (coverage patterns), efficiency (energy/time trade-offs), safety (risk sensitivity), and task (completion criteria).
Graph Reasoning: The LLM performs iterative reasoning over the pruned $G_{global}$ , conditioned on the generated strategy and episode memory. It outputs a sequence of global waypoints ( $\tau_g$ ) representing long-range coverage priorities.

3. Fast-Thinking Module (Local Execution)

This module operates in real-time to execute local motions:

Policy Network: A graph-based RL policy (using attention mechanisms) processes the local belief graph ( $G_{local}$ ). The input includes node positions, utility values (observable frontiers), and a binary "guidepost" signal indicating if a node lies on the global path.
Instruction Following: To ensure the local policy adheres to long-horizon guidance without relying on myopic rewards, the training objective includes a specific reward term. This term penalizes the normalized deviation between the locally selected waypoint and the next global waypoint provided by the slow-thinking module. This encourages the agent to follow the global path while retaining the flexibility to react to immediate local obstacles.

Key Contributions

Fast-Slow Thinking Paradigm: The introduction of a hierarchical framework that separates global semantic reasoning (LLM) from local geometric control (RL), allowing each to operate at its appropriate scale.
Modularity-Based Pruning: A mechanism to reduce the complexity of global graph reasoning by retaining only structurally informative communities, enabling efficient LLM processing of large-scale maps.
Environment-Conditioned Strategy: A method for translating natural language environment descriptions into structured, interpretable exploration strategies that adapt to specific environmental characteristics (e.g., connectivity, obstacle density).
Instruction-Following RL: A novel reward shaping technique that trains the local RL policy to follow global waypoints, effectively bridging the gap between long-term planning and short-term execution.

Experimental Results

The authors evaluated FARE in both simulation and real-world hardware deployments:

Simulation (Gazebo): Experiments were conducted in three environments: Indoor (office), Forest (outdoor with natural obstacles), and Warehouse (dense aisles). FARE was compared against state-of-the-art baselines (TARE, DSVP, ARiADNE, HEADER).
- Performance: FARE achieved the shortest travel distances and fastest completion times in the Forest and Warehouse environments. In the Indoor environment, it performed on par with the best baselines.
- Efficiency: The results indicate that FARE reduces unnecessary backtracking by systematically incorporating global structural cues, unlike baselines that often defer peripheral areas.
Hardware Validation: FARE was deployed on an Agilex Scout-mini wheeled robot equipped with a Jetson AGX Orin and a 3D LiDAR.
- Scale: The system successfully explored a complex, large-scale indoor teaching building (200m × 130m) without manual intervention.
- Robustness: The system maintained stable runtime performance, demonstrating the feasibility of running LLM-based inference onboard for real-time global guidance.

Significance and Claims

The paper claims that FARE represents a significant advancement in autonomous exploration by effectively integrating agent-level semantic reasoning with fast local control. Its primary significance lies in:

Adaptability: The ability to adjust exploration strategies based on high-level environmental descriptions rather than fixed hyperparameters.
Long-Horizon Planning: Overcoming the limitations of sparse rewards in RL and fixed heuristics in conventional planning by explicitly utilizing long-term structural information.
Real-World Viability: Demonstrating that a hybrid LLM-RL architecture can operate effectively on hardware in large, complex, real-world environments, reducing redundant motions and improving overall exploration efficiency.

The authors conclude that this structured interaction between global reasoning and local decision-making enables robots to exploit long-term information and adapt their behavior as the environment unfolds, a capability that existing methods struggle to achieve. Future work is proposed to extend this to multi-robot coordination and incorporate vision-based semantic perception.