BEV-Diffusion: Camera-Only Traversability Navigation with BEV Memory and Efficient Cross-Robot Adaptation

Iana Zhura, Timofei Kozlov, Dmitrii Plotnikov, Ivan Valuev, Miguel Altamirano Cabrera, Dzmitry TsetserukouarXiv 2026 · arXiv

TL;DR

BEV-Diffusion is a camera-only navigation framework that jointly learns bird’s-eye view (BEV) traversability, uncertainty, and path planning within a single network. By combining a frozen DINOv2 backbone with a diffusion-based planner and a metric memory, it achieves robust navigation in complex environments and can adapt to entirely new robot platforms in just over two minutes.

What Problem Does This Paper Address?

Autonomous robot navigation typically relies on expensive LiDAR sensors, pre-built maps, or complex modular pipelines that separate perception from planning. This creates three major hurdles:

  1. Cost and Complexity: LiDAR is expensive and power-hungry.
  2. Robot-Specific Traversability: What is "traversable" depends on the robot. A large wheeled robot might roll over grass easily, while a tiny legged robot might struggle. Most models don't transfer this physical intuition between different bodies.
  3. The "Out of Sight, Out of Mind" Problem: Purely visual policies often "forget" obstacles the moment they leave the camera's field of view, leading to collisions during maneuvers or in tight spaces.

Methods

The authors propose a unified architecture called BEV-Diffusion that operates entirely on monocular RGB frames.

Fig. 2. Overview of BEV-Diffusion. A frozen DINOv2 encodes two RGB frames, which are projected and fused into a polar BEV, where the BEV head predicts the traversability p with its epistemic (u) and aleatoric (σ) uncertainty. A transformer denoiser generates polar waypoints on this BEV, which p, u and σ then refine and truncate. Adaptation retrains only the BEV head. Images and maps are illustrative.
  • Visual Encoding: It uses a frozen DINOv2 backbone to extract features from consecutive RGB frames.
  • Polar BEV Representation: Instead of a standard Cartesian grid, the model lifts features into a polar BEV map. This matches the camera's perspective geometry more naturally, providing higher resolution for nearby areas.
  • Graded Traversability: Unlike binary "hit or miss" maps, a ranking loss trains the model to understand degrees of traversability (e.g., preferring asphalt over grass).
  • Diffusion-Based Planning: A transformer denoiser generates paths as polar waypoints. Because the planner is trained jointly with the map, it can use uncertainty-aware guidance to truncate paths where the visual data becomes unreliable.
  • Metric Memory: At deployment, an ego-motion compensated memory keeps track of obstacles even after they are no longer visible to the camera, allowing the robot to "remember" its surroundings.
  • Efficient Adaptation: To move the model to a new robot, only the "BEV head" (~3% of total parameters) needs retraining, while the heavy-duty backbone and planner remain frozen.

Datasets

The model was trained and evaluated on two very different platforms:

  • Wheeled Quadruped (Agibot D1 Max): A 41kg robot used for outdoor scenes (asphalt, grass, pebbles). The dataset included 4,841 frames.
  • Small Legged Robot (Dogzilla S2): A much smaller robot used for indoor navigation through mazes, arches, and obstacles. This dataset included 4,740 frames.
  • Ground Truth: LiDAR, odometry, and IMU data were used only during training to generate traversability labels; they are not required for inference.

Key Insights and Ideas

  • Joint Learning is Key: By learning the map, the uncertainty, and the path together, the robot doesn't just see a path—it understands how much it can trust that path. If the aleatoric uncertainty (noise in the scene) or epistemic uncertainty (unfamiliarity) is too high, the model cautiously shortens its planned trajectory.
  • Graded vs. Binary: Standard navigation often treats everything as either "obstacle" or "free space." BEV-Diffusion’s ranking loss allows it to distinguish between "good" terrain (asphalt) and "passable but bumpy" terrain (grass), leading to more natural path selection.
  • Polar is Better for Vision: Projecting 2D images to a 3D BEV is mathematically "ill-posed" without depth. Using a polar grid helps mitigate this by aligning the map resolution with the angular resolution of the camera.

Key Results

The performance gains over modular and memory-less baselines are significant:

  • Outdoor Success: In an unseen courtyard, the wheeled robot traveled 120m autonomously, completing 96% of the route without intervention, even in low-light conditions after sunset.
  • Rapid Adaptation: The model adapted from the large wheeled robot to the tiny Dogzilla in just 2.3 minutes of retraining. This adapted model achieved 89% task progress in indoor mazes.
  • Memory Matters: In "trap-like" indoor scenes, the BEV memory allowed the robot to retain 62–66% of out-of-view obstacles, whereas baselines retained less than 3%. This enabled the robot to successfully navigate out of dead ends that baffled other models.
Visual comparison of paths. In Scene 1 (top), the model avoids a wall that baselines crash into. In Scene 2 (middle), it turns in place using remembered free space.

Limitations and Future Work

  • Flat Ground Assumption: The current projection method assumes the ground is flat. While it handles small bumps, it might struggle with steep slopes or stairs.
  • Temporal Calibration: The Bayesian fusion treats consecutive frames as independent, which can lead to a loss of calibration over very long sequences.
  • Future Directions: The authors suggest conditioning the model on IMU data (camera inclination) to handle uneven terrain and extending the "BEV head" to predict step heights for more advanced legged locomotion.

Summary updated

Abstract

Robot navigation requires both an understanding of the scene and an efficient adaptation to new robots, yet current methods rely on expensive LiDAR sensors, long mapping and modular pipelines. We present BEV-Diffusion, a camera-only model that learns a bird's-eye view (BEV) traversability map, its uncertainty and the path in a single network. A frozen DINOv2 backbone encodes consecutive RGB frames, which inverse perspective mapping and Bayesian fusion lift into the BEV, and a ranking loss keeps outdoor traversability graded rather than binary, separating asphalt from grass (median 0.94 vs. 0.75). On this map, a transformer denoiser generates polar waypoints in a few DDIM steps, and uncertainty-aware guidance refines the path and ends it where the map is no longer reliable. At deployment, a metric BEV memory keeps obstacles after they leave the camera view and warm starts each sample from the previous, ego-motion compensated path. Because the planner jointly learns the path, the map, and its uncertainty, it selects routes through terrain it can confidently traverse. In an unseen courtyard, a wheeled-legged robot autonomously traveled 120 m, completing 96% of the route without intervention, compared with 75-88% for the baselines. Even under low-light conditions after sunset, the proposed model completed the route without requiring any intervention. Retraining only the BEV head, about 3% of the parameters, adapts the model to a small legged robot in 2.3 min. Indoors, the memory gets this robot out of maze-like traps: it retains 62-66% of the nearby out-of-view obstacles, against at most 3% for the baselines, and the robot achieves 67-100% task progress, compared with 20-40% for the evaluated baselines.
BibTeX
@misc{zhura2026bev,
  author = {Iana Zhura and Timofei Kozlov and Dmitrii Plotnikov and Ivan Valuev and Miguel Altamirano Cabrera and Dzmitry Tsetserukou},
  title = {{BEV-Diffusion: Camera-Only Traversability Navigation with BEV Memory and Efficient Cross-Robot Adaptation}},
  year = {2026},
  eprint = {2611.24155},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.24155},
}
Figure