Source-linked AI summary
Learning Robust Autonomous Navigation and Locomotion for Wheeled-Legged Robots
Joonho Lee, Marko Bjelonic, Alexander Reske, Lorenz Wellhausen, Takahiro Miki, Marco Hutter
TL;DR
Urban delivery robots need to move efficiently across flat ground while negotiating varied terrain and dynamic obstacles, yet their locomotion, navigation, and system integration remain challenging. The paper develops an integrated hierarchical RL system using model-free RL and privileged learning, then validates it in urban missions. The system completed kilometer-scale autonomous navigation in Zurich and Seville across varied terrains and obstacles, while perception limits constrained demonstrations at the robot’s maximum speed.
Problem
Urban wheeled-legged robots must handle varied terrain, dynamic obstacles, short reaction times, and tightly integrated autonomy for last-mile delivery.
Method
The paper integrates model-free RL and privileged learning for hybrid locomotion with a hierarchical learned navigation controller and large-scale urban path planning.
Results
The integrated system completed kilometer-scale autonomous urban missions in Zurich and Seville, navigating stairs, irregular steps, natural terrain, and pedestrians with minimal human intervention.
Takeaways & Limitations
The results support the feasibility of wheeled-legged robots and hierarchical RL for autonomous navigation across complex urban environments.
Takeaways & Limitations
Limited three-meter forward perception and delayed elevation mapping prevented autonomous demonstrations at the robot hardware’s maximum speed of 6.2 m/s.
Abstract
from arXiv · showhide
Autonomous wheeled-legged robots have the potential to transform logistics systems, improving operational efficiency and adaptability in urban environments. Navigating urban environments, however, poses unique challenges for robots, necessitating innovative solutions for locomotion and navigation. These challenges include the need for adaptive locomotion across varied terrains and the ability to navigate efficiently around complex dynamic obstacles. This work introduces a fully integrated system comprising adaptive locomotion control, mobility-aware local navigation planning, and large-scale path planning within the city. Using model-free reinforcement learning (RL) techniques and privileged learning, we develop a versatile locomotion controller. This controller achieves efficient and robust locomotion over various rough terrains, facilitated by smooth transitions between walking and driving modes. It is tightly integrated with a learned navigation controller through a hierarchical RL framework, enabling effective navigation through challenging terrain and various obstacles at high speed. Our controllers are integrated into a large-scale urban navigation system and validated by autonomous, kilometer-scale navigation missions conducted in Zurich, Switzerland, and Seville, Spain. These missions demonstrate the system's robustness and adaptability, underscoring the importance of integrated control systems in achieving seamless navigation in complex environments. Our findings support the feasibility of wheeled-legged robots and hierarchical RL for autonomous navigation, with implications for last-mile delivery and beyond.
INTRODUCTION
Wheeled-legged robots must combine efficient flat-ground travel with obstacle negotiation, but existing locomotion, navigation, and integration approaches remain limited. This work develops an integrated learning-based system and validates it through urban deployments.
- Motivation: Urban last-mile delivery requires robots that travel efficiently on flat ground while overcoming obstacles such as stairs and uneven terrain.Traditional wheeled robots cannot effectively surmount these obstacles, motivating wheeled-legged platforms.
- Motivation: Wheeled-legged robots can combine high-speed locomotion over moderate surfaces with agility on challenging terrains, but autonomous deployment requires seamless locomotion-navigation integration.The paper identifies hybrid locomotion, smooth navigation, and complete system integration as central challenges.
- Challenges: Existing hybrid locomotion methods use heuristics, predefined gaits, handcrafted patterns, or computationally expensive trajectory optimization, limiting adaptability across situations.Some gait-adaptation results remain limited to indoor or moderate terrains and mostly forward motion.
- Challenges: Navigation planning must account for dynamic hybrid locomotion because urban routes combine fast traversal over open areas with stairs, uneven terrain, and other obstacles.Traditional planners can overlook the robot’s motion characteristics and produce sub-optimal plans.
- Challenges: At multiple meters per second, planners taking several seconds to compute paths require long foresight and may cause collisions in dynamic environments.Human-present environments therefore require faster and more frequent decision-making.
- Approach: The proposed system integrates model-free RL hybrid locomotion control with large-scale navigation, using hierarchical control to coordinate learned navigation and locomotion.The system includes a high-level controller for local navigation and path following, and a low-level controller for joint and wheel commands.
- Validation: Real-world testing in Zurich and Seville demonstrated kilometer-scale autonomous missions through complex terrains and static and dynamic obstacles.The validation included indoor and outdoor urban locations and mock-up delivery missions.
A. Robot
The system combines learned locomotion and mobility-aware navigation with large-scale urban planning, enabling wheeled-legged operation across varied terrain and kilometer-scale missions. Its demonstrations include adaptive gait switching, obstacle negotiation, and 8.3 km autonomous navigation with minimal human intervention.
- Navigation system: Two neural policies coordinate navigation and locomotion, with the high-level controller generating velocity targets and the low-level controller controlling joint actuators.The mobility-aware high-level controller directly computes velocity targets at high frequency and processes the low-level policy’s latent state, terrain heights, and visited positions.
- Urban deployment: 8.3 km of urban navigation was completed in Zurich with minimal human intervention across multiple long-distance experiments and 13 distant goal points.The urban workflow used offline laser scanning, a navigation graph, onboard localization, and shortest-path planning toward GPS goals.
- Locomotion performance: The robot averaged 1.68 m/s with mechanical COT 0.16, achieving three times the speed and 53% lower COT than the compared ANYmal robot.The comparison considered output mechanical power only; heat and transmission losses were not included.
- Obstacle negotiation: The robot navigated around pedestrians and thin obstacles while traversing discrete terrain features such as steps and stairs.These capabilities were demonstrated in populated urban environments, including slopes and stairs.
- Locomotion control: Privileged-learning locomotion switches between driving on flat terrain and stepping on uneven surfaces without handcrafted gait heuristics.The controller was trained with additional motion, terrain, and noiseless exteroceptive information available only during training.
- Operational boundaries: Three missions were stopped proactively because children were in the path or waypoints became untraversable after environmental changes.In the tall-grass case, the robot stopped safely and operators manually triggered global replanning.
E. Human safety
The integrated system combines exploratory, mobility-aware navigation with adaptive locomotion to negotiate blocked routes, narrow spaces, stairs, steps, rough terrain, and pedestrians. It achieved responsive, kilometer-scale urban autonomy while retaining identifiable failure boundaries involving localization, memory capacity, and baseline tracking.
- Obstacle negotiation: The robot explored beyond blocked routes, used positional memory to find alternate paths, and safely navigated narrow passages without collision.It traversed a gap as wide as the robot with two doors and a human standing between them.
- Adaptive traversal: The hierarchical controller adapted gait and route choice to terrain, including finding stairs or a viable approximately 20 cm step.Its traversability assessment was direction-dependent, allowing higher steps during descent than ascent.
- Adaptive traversal: The locomotion controller handled extreme obstacles, including a 60 cm drop and a 40 cm block where all wheels temporarily lost contact.The robot maintained body stability during the drop and used its knees to crawl until wheel contact was regained.
- Limitations: Failures remained when exploration exceeded memory capacity, while localization loss prevented reaching the goal despite safe local operation.A memoryless policy also struggled with repetitive behavior and local minima; baseline failures involved replanning delay and tracking error.
- Responsiveness: The high-level controller processed observations and neural-network inference in 0.34 ms on average, whereas the baseline sometimes required more than a second to update its plan.The comparison used a desktop machine for the baseline and highlights the responsiveness of the learned controller.
- Urban autonomy: The robot completed kilometer-scale autonomous urban missions and navigated stairs, irregular steps, natural terrain, and pedestrians with minimal human intervention.The integrated system combined mobility-aware navigation planning with hybrid locomotion for complex environments.
SUPPLEMENTARY MATERIALS
The supplementary materials describe the sensing, planning, simulation, and learning components supporting autonomous navigation, then evaluate navigation-policy components and training choices.
- Navigation: The system uses LiDAR-based localization against a pre-scanned point-cloud map, with Dijkstra computing shortest paths on an offline navigation graph.Localization combines LiDAR, IMU, and joint-encoder data; the graph is manually designed from reality-capture data.
- Navigation: Anchor pursuit converts sparse global paths into moving sub-waypoints, allowing the hierarchical controller greater freedom to circumnavigate obstacles.The method avoids requiring exact tracking of interpolated waypoints.
- Perception: The robot combines LiDAR terrain mapping with stereo-camera human detection to augment locomotion and navigation scans with safety margins around people.The mapping system provides local elevation maps for foot and base scans, while the camera detects human positions.
- Learning: The navigation policy uses terrain and trajectory-dependent rewards, including a near-goal stability reward active only during high-level policy training.The reward structure separates high-level, low-level, and regularization objectives.
- Evaluation: Ablations show that removing memory reduced distant-goal SPL by approximately 16%, while removing temporal abstraction or modularity made complex navigation harder.The reported component analysis used comparisons across navigation policies and training environments.
- Evaluation: Hierarchical controllers stepped mainly when obstacles required it, unlike a fixed-CPG baseline that continued stepping when unnecessary.The comparison attributes the baseline behavior to the fixed gait-frequency design.
- Evaluation: The terrain fitness function retains terrain parameters when expected traversal scores fall between thresholds, with success defined by velocity-tracking error below 20% of command speed.The score is computed over policy-generated trajectories for each terrain parameter.
- Learning: Privileged learning trains a teacher with simulation-only terrain and dynamics information, then transfers behavior to a student using noisy real-world observations.The student policy uses recurrent temporal features to improve robustness under disturbances and observation noise.