Monocular Navigation Surpasses Depth-Based Systems on R2R-CE

Navigation policies often gain accuracy by depending on depth, multiple cameras, or maps, but those inputs tie a system to a particular sensor suite and robot geometry. Robostral Navigate instead uses a monocular RGB stream and expresses a navigation target in image space. The paper reports that this design reaches 77.4% success on unseen R2R-CE environments, exceeding both the reported best monocular method and the strongest depth- or multi-camera system on that benchmark.

Core Contribution

The central design choice is to separate long-horizon semantic navigation from short-horizon control. An 8B vision-language model selects a waypoint from visual observations and language instructions, rather than predicting robot-frame actions directly. By representing the target through a point in the current image, metric displacement, and arrival orientation, the authors aim to avoid calibration to camera intrinsics, scene scale, and platform-specific coordinate conventions.

That interface is useful because it is neither a purely symbolic subgoal nor a full trajectory. The waypoint gives a downstream controller a geometric target while leaving dense motion generation to a separate policy. Figure 1 depicts this division: the visual-language component operates at 0.5 Hz, whereas the diffusion policy produces trajectories at 10 Hz from the inferred waypoint, its context frame, and the current frame. A motion-tracking controller then converts the trajectory into actuator torques.

Technical Approach

The authors generate 2.4 million trajectories across 350,000 simulated scenes, using simulated data to make the training set much larger than a conventional real-robot collection. At inference, the vision-language model points when a target is visible; when it is outside the field of view, pointing is omitted and the prediction retains displacement and orientation. The diffusion policy fills in the local trajectory between these sparse waypoint decisions.

Training efficiency is another substantive part of the contribution. The paper packs full episodes into a sequence and uses prefix caching, reporting a 22× reduction in training tokens and a reduction in supervised fine-tuning time from months to days. Its tree-based attention mask prevents the model from conditioning on preceding ground-truth actions, an attempt to force action predictions to remain tied to visual evidence rather than exploit the logged trajectory. The authors then apply reinforcement learning to improve exploration and recovery.

The paper also presents cross-robot deployments on the Galaxea R1 and Hiwonder JetAuto. According to Figure 2, these platforms differ in morphology, camera height, camera configuration, and base geometry; the deployments retain the same visual-language and diffusion-policy weights, changing only the low-level controller. This is a meaningful test of the image-space interface, although the paper’s reported benchmark evidence remains the clearer quantitative basis for the claim.

Results and Analysis

On R2R-CE, the post-reinforcement-learning checkpoint reaches 79.43% success in seen environments and 77.43% in unseen environments, compared with 75.96% and 73.40% for the supervised-fine-tuned baseline. The gains are 3.47 and 4.03 percentage points, respectively. The peak validation run reaches 80.46% on seen environments while retaining 77.43% on unseen environments, and the paper reports 74.2% SPL. These results suggest that the RL stage improves recovery without merely making the policy more brittle on held-out scenes.

The headline benchmark comparison is the unseen R2R-CE score: 77.4% success, 10.5 points above the best monocular method and 5.3 points above the strongest depth- or multi-camera system cited by the authors. That is stronger evidence than a hardware-cost argument alone: on this benchmark, the sensor-minimal policy is reported to outperform systems with richer perception inputs.

On the more difficult RxR-CE benchmark, Robostral Navigate reports 75.1% success, 68.7% SPL, and 3.47 m navigation error. Against Qwen-RobotNav-8B, it improves success by 1.7 points and SPL by 5.2 points, from 73.4% success and 63.5% SPL. The smaller success margin on RxR-CE matters: the method’s advantage is clearest in path efficiency, not an overwhelming gap in task completion. Still, the consistency across R2R-CE and RxR-CE supports the paper’s claim that monocular, visually grounded waypoint prediction can remain competitive on long instructions without depth or panoramic sensing.

Scope of the Evidence

The evaluation is concentrated on two continuous-environment navigation benchmarks, so it does not establish performance across the wheeled, legged, and aerial settings named as potential targets. The cross-robot examples demonstrate weight transfer, but the paper does not provide comparable success-rate tables for those physical platforms. The result is therefore best read as strong benchmark evidence for a scalable monocular navigation recipe, rather than a complete demonstration of general-purpose robot navigation.