Autonomous Trail Racing with a Bicycle Robot
Abstract
Navigating off-road trails on a mountain bike requires diverse locomotion skills, including aerial maneuvers, cornering, and slip recovery, paired with sophisticated perception for scene understanding. Translating these capabilities to autonomous, two-wheeled robots remains a formidable challenge. We show that autonomous, high-speed trail navigation can be achieved through a single idea: encoding the trail as a semantic layer of the terrain. In simulation, this layer embeds the navigation task into the training environment; during deployment, it is reconstructed onboard from camera and LiDAR data. First, we introduce a unified learning framework that treats local navigation and locomotion as a joint optimization problem, bypassing the twist or pose command interface of classical control hierarchies. With fixed reward weights, the same framework produces both trail racing and agile parkour, shaped only through terrain semantics and episode length. Second, our perception pipeline combines a frozen vision foundation model with a lightweight dual-head decoder that predicts per-pixel terrain semantics and confidence scores, which are fused with LiDAR range measurements into an egocentric semantic elevation map at 30 Hz onboard an NVIDIA AGX Orin. We validate our framework on the Ultra Mobility Vehicle (UMV), a bicycle robot that autonomously raced across all sections of a kilometer-long, unstructured mountain bike trail at speeds exceeding 5 m/s and averaging 3.26 m/s. In simulation, our policy fails up to an order of magnitude less often than hierarchical baselines while driving faster.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The paper presents a comprehensive framework for autonomous high-speed trail racing on a bicycle robot, achieving state-of-the-art performance through a novel integration of semantic terrain encoding, unified reinforcement learning, and real-time perception.
Methodology
The paper proposes a unified framework for autonomous trail racing on a bicycle robot by encoding the navigable path as a semantic layer within the terrain mesh. This approach eliminates the need for a high-level planner or explicit pose commands, allowing locomotion and local navigation to be learned jointly via reinforcement learning. The perception pipeline (SemMap) is particularly strong, leveraging a frozen DINOv3 backbone with a lightweight dual-head decoder to produce per-pixel segmentation and confidence scores at high frequency (30 Hz on an AGX Orin). The use of feature reconstruction for confidence estimation is a clever, lightweight alternative to complex uncertainty quantification methods. The integration of this semantic map with LiDAR data to form an egocentric semantic elevation map provides a robust state representation for the control policy.
Results & evaluation
The experimental validation is extensive, covering both simulation and real-world deployment. The authors demonstrate the system's capability on a 1km mountain bike trail, achieving average speeds of 3.26 m/s and peaks over 5 m/s. They provide rigorous comparisons against hierarchical baselines in simulation, showing an order-of-magnitude improvement in failure rates. The inclusion of parkour tasks demonstrates the versatility of the framework. The real-world results are compelling, showing that the robot can handle diverse obstacles like roots, rocks, and drops autonomously.
Reproducibility
The authors commit to open-sourcing their simulation and perception tools (TL and SemMap), which significantly enhances reproducibility. The paper provides detailed descriptions of the hardware setup, software stack (ROS2, TensorRT), and training procedures. The specific parameters for the reward functions and terrain generation are described, although some details are deferred to the appendix. The use of standard tools like Isaac Lab and PyTorch further aids reproducibility.
Limitations
The system relies on a high-fidelity digital twin for training specific parkour maneuvers, which may not generalize perfectly to unseen real-world variations without fine-tuning. The perception pipeline, while fast, still requires significant computational resources (AGX Orin), limiting its applicability to lower-power platforms. The paper focuses on a specific robot (UMV), and while the framework is claimed to be platform-agnostic, the transfer to other embodiments is only briefly mentioned in simulation.
Broader impact
This work has significant implications for the field of autonomous robotics, particularly for statically unstable platforms like bicycles and motorcycles. The approach of embedding navigation tasks into the terrain semantics could inspire similar solutions in other domains where explicit planning is difficult. The high-speed autonomous traversal of unstructured terrain is a major step forward for outdoor robotics. The paper presents a comprehensive framework for autonomous high-speed trail racing on a bicycle robot, achieving state-of-the-art performance through a novel integration of semantic terrain encoding, unified reinforcement learning, and real-time perception.
Author of this paper? Send a correction.