← Paper Digest

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 Tsetserukou

November 2026

TL;DR A camera-only navigation model that predicts a graded bird's-eye-view traversability map, its uncertainty and a diffusion-generated path in one network, and ends the path where the map becomes unreliable. On a wheeled quadruped in an unseen courtyard it covered 96% of the route without intervention (baselines 75-88%), and retraining 3% of its parameters for 2.3 min moved it to a small legged robot.

What they do

Navigation for legged robots from a single camera, with no prior map or localisation. One network predicts a top-down (BEV) traversability map with its uncertainty and generates a path on it by diffusion. LiDAR, odometry and IMU are used only to make training labels.

How it works

  • BEV from two frames. A frozen DINOv2 with small adapters encodes two RGB frames 0.1 s apart; the features are projected onto a flat ground plane and fused by adding log-odds into a polar grid. A head predicts graded traversability (0 impassable to 1 free) with epistemic and aleatoric uncertainty, and a ranking loss keeps asphalt and grass apart.
  • Path diffusion. A 26M-parameter transformer denoiser generates 50 polar waypoints in five DDIM steps. Traversability and uncertainty gradients refine the path, which is cut where confidence falls below 0.5. A plan takes 42 ms on an RTX 5090.
  • BEV memory. At deployment, an ego-centric grid shifted by ego-motion estimated from the images keeps obstacles that left the view, warm-starts each sample and turns the robot towards remembered free space when it is stuck.
  • Adaptation. A new robot retrains only the BEV head, 1.1M of 40M trained parameters.

Results

  • On a 41 kg Agibot D1 Max wheeled quadruped, over 120 m in an unseen courtyard, 96% of the route passed without intervention versus 75-88% for BEVFormer + A*, BEVFormer + diffusion and TG-Diff, with no intervention after sunset.
  • The learned BEV detects 69.5% of blocked cells, against 5.6% for image-space traversability projected onto the ground.
  • Adapting to the small Dogzilla S2 took 2.3 min instead of 19.1 min from scratch, for 89% versus 96% indoor task progress.
  • In the round room, the memory keeps 62-63% of out-of-view obstacles versus 2-3% for the BEVFormer baselines; across the two trap-like scenes progress is 67-100% versus 20-40%.

Limitations

The flat-ground projection limits it to mostly flat terrain. In unfamiliar scenes the map detects fewer obstacles: blocked recall in the unseen courtyard is 37-40%, below the BEVFormer baselines’ 55-69%. Fusion treats frames as independent, so more frames hurt calibration.

Figures

From Figure 1: Top, on the wheeled quadruped the model reconstructs graded BEV traversability from the camera and gen...
From Figure 1: Top, on the wheeled quadruped the model reconstructs graded BEV traversability from the camera and generates a path on it by polar diffusion. Bottom, retraining only the BEV head (1.1M parameters, about 3% of the trained ones) adapts it in 2.3 min to the small Dogzilla S2 quadruped, which plans through an arch.
From Figure 2: A frozen DINOv2 encodes two RGB frames, which are projected and fused into a polar BEV where the BEV h...
From Figure 2: A frozen DINOv2 encodes two RGB frames, which are projected and fused into a polar BEV where the BEV head predicts traversability with its epistemic and aleatoric uncertainty. A transformer denoiser generates polar waypoints on this BEV, which traversability and uncertainty then refine and truncate. Adaptation retrains only the BEV head.
From Figure 3: Wheeled quadruped outdoors, Scene 1: camera view (left) and the predicted BEV map (right, 0.3-6 m) wit...
From Figure 3: Wheeled quadruped outdoors, Scene 1: camera view (left) and the predicted BEV map (right, 0.3-6 m) with the paths of BEV-Diffusion, A*, a diffusion planner and TG-Diff; orange marks high uncertainty. All baseline paths enter the LiDAR-labelled door wall within their first 2 m, while BEV-Diffusion turns towards the open passage.
From Figure 4: Small quadruped indoors, Scene 2 (round room): camera view (left) and BEV map (right) with the paths o...
From Figure 4: Small quadruped indoors, Scene 2 (round room): camera view (left) and BEV map (right) with the paths of BEV-Diffusion, A*, a diffusion planner and TG-Diff. Below the dashed line lies the camera's blind zone, where the map shows the BEV memory (teal where the floor was never seen). Only BEV-Diffusion keeps clear of the obstacles, turning in place towards free space kept in memory.
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.

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.