VISTA: Scale-Aware Visual Navigation via Action History Conditioning

Figure 6: VISTA takes as input a sequence of image observations O_{t} and a goal image o_{g} at timestep t . The DINOv3 encoder encodes the observation sequence and the goal observation, respectively. The resulting patch tokens are then processed by the projector network, matching feature dimension while downsampling them by 2x. The backbone is composed of 6 transformer layers with 512 dimensions, 8 attention heads, and a feed-forward dimension of 4\times 512 . Our proposed architecture outputs waypoints \hat{w}_{1:T} , alongside metric distance \hat{d} .

Figure 17: Loop robustness for topological navigation. For a given topological map corresponding to a loop, each baseline is tasked to traverse the loop 5 times. Accumulated distance prediction errors hinder sub-goal selection, leading to collision or getting stuck.

Figure 1: Deployment time-lapse of VISTA: Scale-Aware Visual Navigation via Action History Conditioning, achieves zero-shot deployment in unseen environments by conditioning on action history to improve scaling across robotic platforms and trajectories.

Figure 3: Quantitative results: Path length and topological map (topomap) crossed, across scenes.

Figure 4: Visual of the loop experiments. VISTA reliably achieves 5 loops across 5 trials.

Figure 5: Dynamic perturbation. Our model navigates the Office-Lab environment using a topological map built under nominal conditions ( door open ). At deployment time, the door is closed, then reopened. Across all 5 trials , this halt is observed.

Figure 7: Dataset environment composition. Indoor-outdoor split per dataset ( left ); SACSoN , Scand , Recon , Tartan Drive and Go Stanford . Per-environment breakdown across indoor and outdoor settings ( right ); where the two smaller slices are Gymnasium and Library .

Figure 8: Object encounter probabilities per trajectory across datasets: How likely am I to encounter this object when sampling a trajectory from this dataset?

Figure 13: ViNT , NoMaD , MetricNet , VISTA and the VISTA w/o AH for all (3 trials) outdoor deployment results with the Reference trajectory. Visualization is qualitative as trajectory-to-map alignment is approximate. See Table 2 and Figure 3 for qualitative results.

Figure 16: Visual representation of the loop experiments with VISTA , ViNT , VISTA w/o AH , NoMaD and MetricNet with Reference path. For visibility, we report one trial for ViNT . We highlight collisions with red stars. Note that the figure is qualitative as trajectory-to-map alignment is approximate. For quantitative results, see Figure 17 .