Autonomous Trail Racing with a Bicycle Robot

Fabian Jenelten, Deepana Ishtaweera, Pablo Garcia, Liana Margolese, Igor Bogoslavskyi, Luca Buonocore, Jan Migliazza, Aleksandrs Ecins, Annika Wollschlaeger, Dario Bellicoso, Gabriel Nelson, Abel Gawel, Marco Hutter

November 2026

The robot racing autonomously down a forest section of the trail.

TL;DR A bicycle robot learns trail racing and parkour from one RL setup in which the trail is a semantic layer of the terrain, rebuilt onboard from camera and LiDAR at 30 Hz. It raced every section of a kilometer-long mountain-bike trail at 3.26 m/s on average, peaking above 5 m/s.

What they do

Racing a two-wheeled, statically unstable robot down a mountain-bike trail couples perception and agile control. Instead of a planner sending velocity commands to a locomotion policy, the authors encode the trail as a semantic layer of the terrain: painted into the mesh in simulation, rebuilt from camera and LiDAR on the robot. A single RL policy then does local navigation and locomotion together on the UMV bicycle robot.

How it works

Training. The policy sees proprioception and a two-layer height scan (elevation plus an on-trail flag), with no goal position or twist command. Rewards are forward progress, staying on the trail and a track-stand in the final zone; standing pays more per step than driving, so arriving early raises the return. Weights stay fixed across tasks, and behavior is shaped through the terrain and the episode length.
Terrain. 1000 procedurally generated trail patches (over 46 km), plus scanned and hand-annotated digital twins of parkour tracks and the target trail.
Perception (SemMap). Frozen DINOv3-s features, a learned upsampler and a small decoder give per-pixel classes and a confidence from feature reconstruction; confidence-weighted fusion with LiDAR builds a 15 x 15 m egocentric map at 30 Hz on an NVIDIA AGX Orin.

Results

Real trail: six runs, picked from 79, together covered all sections, 946 m at 3.0-3.8 m/s average per run and 5.2 m/s peak, 2.3x faster than the earlier teleoperated UMV on the upper section.
Indoors: 130 continuous laps at 3.54 m/s (6.85 s per lap vs 6.1 s in simulation); MoCap-assisted parkour succeeded 8/10 vs 99% in simulation.
In simulation the policy survived 91-99% of runs at 3.3 m/s, against 46-85% at 2.2 m/s for the best hierarchical baseline.
With 20 labeled images the decoder reaches 33.9% mIoU on RELLIS-3D vs about 24% for fine-tuned DeepLabV3+ and SegFormer-B0.

Limitations

Outdoor runs are short: the median run covered about 74 m and the longest lasted 76 s. Failures came from leaving the trail (23/67), wheel slip on gravel (20/67), roots, skinnies or thin trunks too small for the height scan (14/67) and mapping errors from trees and tall grass (10/67). The policy has no global routing at forks.

Videos

Indoor parkour: slalom, ramp jump over a gap and staircase landing.

Trail segmentation overlaid on camera footage of the forest trail.

Figures

From Figure 1: The UMV autonomously riding a mountain-bike trail in Zurich, cornering a curve (left) and crossing pum... — From Figure 1: The UMV autonomously riding a mountain-bike trail in Zurich, cornering a curve (left) and crossing pump-track waves (right).

From Figure 3: Approach overview. (A) The policy observes proprioception, an operator-triggered stand command and a s... — From Figure 3: Approach overview. (A) The policy observes proprioception, an operator-triggered stand command and a semantic height scan, and outputs joint targets at 50 Hz for the rear wheel, the fork and the two torso joints. (B) Training on synthetic trails or digital twins, for both racing and parkour. (C) Semantic pipeline: frozen DINOv3 features and upsampler, a light decoder for per-pixel segmentation, a reconstruction head whose cosine similarity gives per-pixel confidence, and fusion with odometry and point clouds into an elevation map at 30 Hz.

From Figure 9: The kilometer-long Allmend trail in Zurich (red), cut in two by a pedestrian path (dotted blue), with ... — From Figure 9: The kilometer-long Allmend trail in Zurich (red), cut in two by a pedestrian path (dotted blue), with the skill elements the UMV crosses: a rock bed, seven waves, skinnies, tabletop drops, a wall-ride, a log, a seesaw ramp, tight and overgrown curves and a steep descent.

From Figure 7: (A) MoCap-assisted indoor parkour with a slalom, a lift-off ramp and a landing staircase. (B) Autonomo... — From Figure 7: (A) MoCap-assisted indoor parkour with a slalom, a lift-off ramp and a landing staircase. (B) Autonomous indoor racing with onboard compute and perception: 130 consecutive laps in 891 s. (C) Base velocity per lap phase and position traces of all laps colored by speed. (D) The map with confidence-weighted fusion (top) and without it (bottom), where low-confidence views in fast turns corrupt it.

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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong