Vision-and-language navigation (VLN) research typically evaluates long-horizon tasks where instructions are assumed to be physically executable. The authors of PASS-VLN argue that this assumption leaves a critical gap: body-dependent local execution, such as whether a robot can pass through a gap and execute a command without collision. To isolate this problem from planning and memory, their benchmark fixes the scene, instruction, start pose, and reference route while varying the robot's body width. The robot must complete the task when the passage is feasible and stop before blockage when it is not.
The proposed model encodes body width by inflating obstacles in a depth-derived local bird's-eye view, then fuses this geometric information with visual and linguistic context. It predicts discrete latent action codes for continuous closed-loop execution. In experiments, PASS-VLN improves collision-free success by roughly 12 percentage points over a strong VLN model fine-tuned on the benchmark, while adapting its reach-versus-stop behavior to body constraints and generalizing to unseen body widths.
When integrated with a high-level planner, PASS-VLN achieves higher success with fewer local commands, and real-world deployment demonstrates its usefulness as a low-level executor. The benchmark and model are planned for release. As this digest draws on a single source, the results are as reported by the authors rather than independently verified.