Source-linked AI summary
From Perception to Decision: A Data-driven Approach to End-to-end Motion Planning for Autonomous Ground Robots
Mark Pfeiffer, Michael Schaeuble, Juan Nieto, Roland Siegwart, Cesar Cadena
TL;DR
The paper addresses how to reduce hand-tuning in robot navigation by learning steering directly from laser and target data. It trains a CNN on simulated expert demonstrations and reports transfer to unseen virtual and real environments, while identifying practical scope limits.
Problem
Robot navigation traditionally requires hand-tuning and decoupled preprocessing, sensing, object detection, and planning stages.
Method
A CNN learns the end-to-end mapping from simulated expert demonstrations of laser measurements and target information to steering commands.
Results
The learned model transfers navigation knowledge from training environments to unseen complex environments and can be trained in simulation for deployment on a real platform.
Takeaways & Limitations
Target-oriented navigation and collision avoidance are achievable with local information in maze-like environments, including on a real robotic platform.
Takeaways & Limitations
The approach remains a local planner that relies on a global planner for targets and is limited in wide open spaces with glass or clutter.
Abstract
from arXiv · showhide
Learning from demonstration for motion planning is an ongoing research topic. In this paper we present a model that is able to learn the complex mapping from raw 2D-laser range findings and a target position to the required steering commands for the robot. To our best knowledge, this work presents the first approach that learns a target-oriented end-to-end navigation model for a robotic platform. The supervised model training is based on expert demonstrations generated in simulation with an existing motion planner. We demonstrate that the learned navigation model is directly transferable to previously unseen virtual and, more interestingly, real-world environments. It can safely navigate the robot through obstacle-cluttered environments to reach the provided targets. We present an extensive qualitative and quantitative evaluation of the neural network-based motion planner, and compare it to a grid-based global approach, both in simulation and in real-world experiments.
I. INTRODUCTION
The paper proposes learning end-to-end navigation from expert demonstrations to reduce hand-tuning and integrate sensing with steering. The approach uses local laser and target information, transfers to unseen environments, and is evaluated in simulation and on a real robot.
- Expert demonstrations are generated in simulation with an existing motion planner to teach navigation behavior and collision avoidance.
- The robot navigates using sensor data and a relative target position without requiring a global map.
- The learned model is tested in simulation and on a real robotic platform, including unknown environments more challenging than the training environment.
- The approach learns a data-driven end-to-end motion planner from laser range findings to motion commands.
- The work includes deployment on a real robotic platform and extensive comparison with a global-map motion planner.
II. RELATED WORK
Prior work shows that learning can reduce hand-tuning and support end-to-end control, but existing mobile-robot approaches lack some combination of target orientation, full motion control, or environment generality.
- Learning-based navigation has reduced hand-tuning, but some methods require maps or road networks and remain specific to a single environment.
- End-to-end learning has been applied to perception-control coordination, including robotic-arm torque planning with convolutional neural networks.
- A UAV controller learned left-right steering from images, while forward motion remained under human control.
- A laser-based auto-encoder approach demonstrated collision avoidance but did not use target positions, preventing target-directed local planning.
III. APPROACH
The paper formulates navigation as directly mapping sensor and target observations to steering commands. Expert demonstrations supervise a CNN that replaces separately decoupled perception and decision stages.
- Robots must extract sensor information, model observation-action relationships, and make decisions promptly during deployment.
- The proposed approach directly computes steering commands from sensor and target data instead of solving these tasks independently.
- A parameterized function maps sensor vector y and goal information g to steering commands u, trained against expert commands u_exp.
B. End-to-end model
The model processes 2D laser measurements with convolutional residual blocks, fuses the resulting features with relative target information, and predicts steering commands through fully connected layers.
- The CNN processes laser data before combining extracted features with relative target information in fully connected layers.
- The convolutional component uses two residual building blocks with shortcut connections to reduce training complexity.
- Two model variants differ in fully connected layer widths, while sharing the same convolutional component.
C. Model training
The model learns expert steering behavior from simulated laser, target, and velocity-command demonstrations using supervised neural-network training. Its design also targets predictable inference without external laser preprocessing or temporal memory.
- Data and optimization: Training tuples combine laser measurements, relative target information, and expert translational and rotational velocity commands.Tuples are randomized before mini-batch training with Adam optimization.
- Data and optimization: The loss compares predicted and expert velocity commands across mini-batches and is optimized by backpropagation.The mini-batch contains multiple training tuples, and gradients are computed with respect to model parameters.
- Inference: The neural network computes steering commands directly from laser data without external preprocessing, keeping query complexity independent of environmental complexity.The paper identifies predictable query time as an advantage over multi-stage approaches.
- Inference: Steering commands are computed frame-by-frame without internal or external memory of previous inputs or outputs.
IV. EXPERIMENTS
The experiments evaluate a ROS expert planner and a learned deep planner across simulation and real-robot settings. Training uses simulated navigation data generated with a layered ROS navigation stack and randomly selected collision-free targets.
- Experimental setup: The evaluation comprises four experiments: two in simulation and two on a real robotic platform.The TurtleBot platform is used both as a simulated model and for real-world tests.
- Robot platform: The real platform is a Kobuki-based TurtleBot equipped with a 270° Hokuyo laser scanner providing 1080 measurements.
- Expert planner: The ROS expert combines grid-based Dijkstra global planning with dynamic-window local planning in the Stage 2D simulator.
- Training data: Training data records simulated laser measurements, relative target positions, and expert steering commands for randomly selected collision-free targets on a 10 m × 10 m map.The train map is used for training-data generation, while additional map data supports real-world testing.
- Training data: Simulation-only training provides deterministic data, avoids noisy sensor effects in basic experiments, accelerates data generation, and enables simulation-to-reality testing.
- Evaluation protocol: Figure 3 reports frame-by-frame error statistics on evaluation data that was not used during training, across three evaluation maps.
C. Frame-by-frame evaluation
Frame-level and trajectory-level simulation evaluations test whether the learned planner reproduces expert behavior and transfers it across maps. Errors are lowest on the training map and increase as map structure differs more substantially.
- C. Frame-by-frame evaluation: The frame-by-frame evaluation compares deep and expert translational and rotational steering commands on previously unseen input/output tuples.The model was trained only on samples from the train map.
- C. Frame-by-frame evaluation: The smallest evaluation error occurs on the train map, increases on eval1, and increases further for both translational and rotational commands on eval2.Eval1 has a different structure but similarly shaped obstacles, whereas eval2 differs more strongly from training.
- C. Frame-by-frame evaluation: The authors attribute larger steering-command deviations to weaker scene understanding when environmental structure differs substantially between train and eval2.
- D. Trajectory comparison in a simulated environment: Trajectory evaluation compares final goal distance, translational energy, travelled distance, rotational energy, and travel time using metrics independent of the ROS planner’s cost function.
- D. Trajectory comparison in a simulated environment: On the train map, deep-planner trajectories are congruent with expert trajectories in most cases and slightly outperform the expert on the selected metrics.The authors state that neither planner was trained or tuned for these metrics.
- D. Trajectory comparison in a simulated environment: On eval1, the deep planner transfers expert-like routing despite differing local topologies, but corrections increase rotational energy while distance and time remain similar.The planner sometimes cuts edges more sharply than the ROS planner.
- D. Trajectory comparison in a simulated environment: Figure 4 contrasts train and eval1 trajectories and relative errors for five metrics, with green denoting ROS-better relative error and black denoting deep-planner-better relative error.
E. Real-world navigation
Real-robot tests assess transfer, repeatability, and manual intervention in cluttered environments. The deep planner drives most missions autonomously, but dense clutter remains a practical boundary requiring operator assistance.
- E. Real-world navigation: Real-world tests use only local laser and target information for the deep planner, while the ROS expert is localized with a prerecorded map.The robot traverses mazes, obstructed corridors, cluttered areas, and long corridors.
- E. Real-world navigation: The real-world model is trained with train and eval2 data to combine basic navigation principles with object shapes likely to appear in reality.Increasing the fully connected layer size improves real-world navigation but not simulation performance.
- E. Real-world navigation: Table I measures manual joystick control as a percentage of total distance traveled during successful navigation.
- E. Real-world navigation: Figure 5 compares one ROS trajectory with six CNN bigFC trajectories on the real robot and includes environmental views.
- E. Real-world navigation: The deep planner completes the majority of missions autonomously, while larger fully connected layers significantly reduce joystick interventions.The authors relate the improvement to handling more diverse features in richer environments.
- E. Real-world navigation: Across six CNN bigFC drives, deep-planner navigation characteristics remain similar and consistent, with wider or later reactions in a few areas.Examples include a wider turn between targets 4 and 5 and a later reaction to tables between targets 6 and 7.
- E. Real-world navigation: The robot could not traverse the cluttered area between targets 10 and 11 fully autonomously because training used relatively well-arranged closed environments.
- E. Real-world navigation: Joystick interventions generally unstuck the robot rather than preventing imminent collisions, and no unstable motion was observed.Unhandled situations more often caused the robot to stop than to follow unpredictable paths.
F. Reaction to sudden changes
The deep planner reacts to a suddenly appearing obstacle by deviating from its route, then corrects course after the obstacle is removed. Figure 6 visualizes the resulting path, sensor inputs, and target position.
- The robot swerves right when a suddenly appearing object blocks its corridor path to the target.
- After the obstacle is removed, the robot corrects its path to approach the target.
- Figure 6 shows the constant-velocity path and setup in the top row, laser findings and relative target position in the middle row, and robot-view images in the last row.The robot-view images are used only for visualization.
V. DISCUSSION
The study demonstrates that a CNN can learn navigation from expert demonstrations and transfer it from simulation to unseen environments and a real robot. The approach remains bounded by local-planning and training-data limitations.
- A CNN learns an end-to-end mapping from raw sensor data to steering commands using an expert operator’s navigation policy.
- The learned navigation model transfers knowledge from training environments to unseen and complex environments.
- The approach can train in simulation and deploy on a real platform while retaining satisfactory navigation performance.The authors identify this as valuable where data generation is expensive.
- The deep planner cannot completely replace a map-based path planner because it remains a local planner relying on global-planner targets in complex environments.
- Wide open spaces with substantial glass or clutter limit the current approach, potentially because training used purely perfect simulation data.
- The planner cannot free itself after entering a convex dead-end region, and its heading sometimes fluctuates before obstacle avoidance.The authors suggest recurrent neural networks with internal memory as a possible remedy.
VI. CONCLUSION
The paper presents a data-driven end-to-end planner that converts local laser findings and a relative target position into steering commands. A CNN learns navigation strategies from expert demonstrations and transfers them across environments, including deployment on a real robot.
- The approach computes steering commands for a differential-drive platform from local laser range findings and a relative target position.
- The CNN learns navigation strategies from an expert operator and transfers this knowledge between different environments.
- Simulation data can train the navigation model for deployment on a real robotic platform in an unseen environment.