💻 computer science

Learning-Based Navigation for Indoor Mobile Robots

This paper presents a learning-based navigation framework for indoor mobile robots that integrates a supervised neural global planner trained on A* trajectories with a local planner combining behavior cloning and PPO-refined Dynamic Window Approach, demonstrating effective and safe goal-directed navigation in both simulated and real-world environments.

Original authors: Tri-Tin Nguyen, Tien-Dat Nguyen, Gia-Uy Le, Vinh Nguyen, Vinh-Hao Nguyen

Published 2026-06-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tri-Tin Nguyen, Tien-Dat Nguyen, Gia-Uy Le, Vinh Nguyen, Vinh-Hao Nguyen

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 trying to guide a robot through a busy, cluttered indoor hallway to get from the front door to a specific office. This paper presents a new "brain" for that robot, designed to make the journey smoother and smarter than the old ways of doing it.

The authors split the robot's brain into two distinct roles, working together like a Navigator and a Driver.

1. The Navigator (The Global Planner)

The Old Way: Traditionally, robots use a strict, rule-based map search (like a very careful but slow GPS) to draw a line from point A to point B. It works, but it can be rigid.
The New Way: The authors trained a "Neural Navigator" by watching an expert.

The Analogy: Imagine a student learning to navigate a city. Instead of memorizing a rulebook, the student watches a master taxi driver (the "Cost-Aware A* Expert") take the best routes. The student learns to predict the next turn just by looking at the map and the destination.
How it works: The robot looks at a digital map of obstacles and learns to predict the next step in the path. It's like a chess player who has studied thousands of games and can instantly "see" the best move without calculating every single possibility from scratch.

2. The Driver (The Local Planner)

The Old Way: Once the Navigator draws the line, a "Driver" (called DWA) has to actually move the robot. The traditional driver is very cautious. It constantly checks: "Is there a wall? Is the goal ahead? Am I moving too fast?" It picks the safest, most conservative move from a list of options. In a crowded room, this can make the robot move jerkily or get stuck.
The New Way: The authors created a "Learning-Based Driver."

The Analogy: Think of the traditional driver as a nervous learner who only picks moves that are 100% safe. The new driver is like a skilled stunt driver who has been trained in two stages:
1. Imitation (Behavior Cloning): First, the driver watches the nervous expert driver and copies their moves exactly.
2. Refinement (PPO): Then, the driver practices in a simulator. It gets a "reward" for moving smoothly and a "penalty" for crashing. It learns to tweak the expert's moves to be smoother and more efficient, but it is strictly forbidden from picking a move that would cause a crash.
The Safety Net: Crucially, this driver doesn't invent new, crazy moves. It only chooses from a pre-approved list of safe steering and speed combinations (the "DWA action lattice"). It's like a driver who can choose from a menu of safe maneuvers but learns to pick the best one for the moment.

The Results: What Happened?

The team tested this system in both computer simulations and on a real robot with wheels and sensors.

Smoother Rides: The new "Learning-Based Driver" moved much more smoothly. If you were riding in the robot, you wouldn't feel the jerky stops and starts of the old system. The "jerk" (sudden changes in speed) was significantly reduced.
Better Paths: In cluttered rooms, the new system found paths that were slightly shorter and avoided obstacles more intelligently than the old rule-based system.
The Trade-off: There was one catch. Because the new driver was so focused on being smooth and safe, it sometimes took a tiny bit longer to reach the destination than the old, frantic driver. The old driver was faster but bumpier; the new driver was a luxury sedan ride—slightly slower, but much more comfortable.

In Summary

The paper shows that by combining a smart, learning-based map reader with a refined, safety-conscious driver, robots can navigate indoor spaces more effectively. They don't just avoid crashing; they move with a human-like smoothness, learning from experts and then practicing to get better, all while staying within the safety rules of the road.

Technical Summary: Learning-Based Navigation for Indoor Mobile Robots

Problem Statement

Indoor mobile robot navigation typically relies on a hierarchical architecture comprising a global planner for map-level route generation and a local planner for real-time obstacle avoidance. While classical methods like A* and the Dynamic Window Approach (DWA) are widely used, they often rely on handcrafted search rules and cost functions. Consequently, these planners can produce conservative motions or struggle with efficiency in cluttered environments. The paper addresses the need for a navigation framework that improves path quality and motion smoothness while maintaining the feasibility and safety constraints inherent in real-world robotic systems.

Methodology

The proposed framework integrates a supervised neural global planner with a learning-based local planner that operates within the structured action space of DWA.

1. Global Path Planning Module

The global planner is designed to generate feasible map-level routes on known indoor costmaps.

Training Data: The planner is trained via supervised learning using trajectories generated by a cost-aware A* expert. This expert planner expands 8-connected neighbors on a 2D costmap, minimizing a cost function $f(n) = g(n) + h(n)$ that includes penalties for steps, turning, and high-cost cells.
Network Architecture: The planner utilizes a convolutional policy network ( $\pi^G_\theta$ ) that takes a five-channel tensor input (encoding obstacle occupancy, costmap values, unknown-space masks, and robot/goal positions).
Output: The network predicts the next-step action as one of eight discrete directions (N, NE, E, SE, S, SW, W, NW).
Training Objective: The model is trained to minimize a combined loss function consisting of an action-classification loss (cross-entropy against expert actions) and an auxiliary path mask loss to provide dense spatial supervision.
Deployment: At runtime, the planner operates autoregressively in grid space, predicting actions, masking invalid moves, and updating the state until the goal is reached or a horizon limit is hit.

2. Local Path Planning Module (Learning-Based DWA)

The local planner generates real-time velocity commands $(v, \omega)$ for path tracking and obstacle avoidance. Unlike end-to-end approaches that predict continuous velocities directly, this method performs discrete candidate selection over the standard DWA action lattice.

Observation Space: The input includes a global context vector (goal position, heading error, velocities, distances) and candidate-wise features for $N=63$ DWA candidates (velocity, costs, and validity flags).
Two-Stage Training:
1. Behavior Cloning (BC): The policy is initialized by imitating the decisions of a conventional DWA expert. The network outputs logits for the $N$ candidates, and a validity-aware logit adjustment is applied during inference to mask infeasible candidates.
2. PPO Refinement: The BC-initialized policy is further refined using Proximal Policy Optimization (PPO). The reward function promotes progress, safety, and smoothness, with penalties for collisions and timeouts. Crucially, the PPO refinement retains the validity-aware masking, ensuring the policy remains constrained to the feasible DWA candidate space.

Key Contributions

The paper outlines three primary contributions:

Hybrid Framework: A deployable navigation system integrating a supervised neural global planner with a DWA-based local candidate selector, preserving the constraint-aware structure of DWA.
Two-Stage Local Policy: A local planning strategy trained first by behavior cloning and then refined by PPO, utilizing validity-aware action masking to ensure safety.
Empirical Validation: Comprehensive evaluation in both simulated and real-world indoor environments demonstrating improved path quality and motion smoothness compared to conventional DWA.

Experimental Results

The framework was evaluated on a differential-drive robot equipped with a 2D LiDAR and an onboard computer, using both Gazebo simulations and a real-world indoor environment.

Global Planner Performance:
- In offline validation, the best checkpoint achieved a rollout success rate of 91.94%.
- Runtime comparisons showed that GPU-based Neural A* (approx. 80ms) is significantly faster than CPU inference (approx. 12.6s) and competitive with classical A* (approx. 68ms).
- In obstacle-rich environments, the Neural A* achieved a 67% path win rate compared to 33% for classical A*, indicating superior path quality in complex settings.
Local Planner Performance:
- Static Scenarios: The Learning-Based DWA reduced path tracking RMSE from 0.1044m (conventional DWA) to 0.0633m and significantly lowered linear jerk (0.8230 vs. 1.7673), resulting in smoother motion. However, it was slightly slower (36.0s vs. 31.8s to goal).
- Obstacle Scenarios: In cluttered environments, the learned policy demonstrated more anticipative avoidance, achieving a shorter path length (8.12m vs. 8.39m) and smoother angular motion (lower RMS angular jerk) compared to conventional DWA.
- Simulation vs. Real-World: The system demonstrated 100% success rate in simulation and was successfully deployed on a real robot, navigating indoor environments with feasible global routes and reliable local commands.

Significance and Claims

The authors claim that the proposed framework effectively integrates learning-based global planning with reinforcement-learning-refined local control. The primary significance lies in the ability to generate feasible global routes and reliable local motion commands that are safer and smoother than conventional methods, particularly in the presence of obstacles.

The paper emphasizes that by selecting actions from the DWA action lattice rather than learning unconstrained continuous commands, the system preserves the feasibility guarantees of DWA while leveraging machine learning to improve adaptability and motion quality. The authors note that while the learning-based approach yields smoother trajectories and better path adherence, it can be more conservative in terms of traversal time compared to the rule-based baseline. The work provides initial evidence of generalization across simulated maps and successful transfer to a real-world robot, though the authors acknowledge that further evaluation under diverse layouts, dynamic obstacles, and sensor noise is necessary.