Successful but Unsafe: An Analytical Benchmark for Revealing Collisions in Vision-and-Language Navigation
We introduce VLN-Safe, a benchmark for analyzing collision safety under two complementary evaluation regimes: sliding, reflecting common VLN evaluation practice, and blocking, where collision-inducing actions prevent forward motion and more directly expose their consequences.
ProofPaper ↗
Key points
- Vision-and-language navigation (VLN) agents are typically evaluated by whether they reach a target and how efficiently they follow an instructed route, yet these metrics provide little information about collision behavior.
- Moreover, continuous VLN evaluation commonly uses a sliding regime, where an agent can slide along obstacles after a collision, potentially masking the effect of unsafe actions on navigation performance.
- Motivated by these findings, we introduce a training-free collision guard that vetoes risky forward actions, executes an escape maneuver, and returns control to the original policy.
- These results reveal a substantial gap between task success and collision-safe navigation and show that lightweight collision-aware intervention can improve both safety and navigation performance.
Sources (1)
- [1]Successful but Unsafe: An Analytical Benchmark for Revealing Collisions in Vision-and-Language NavigationarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 10, 07:53 AM
We introduce VLN-Safe, a benchmark for analyzing collision safety under two complementary evaluation regimes: sliding, reflecting common VLN evaluation practice, and blocking, where collision-inducing actions prevent forward motion and more directly expose their consequences.
Vision-and-language navigation (VLN) agents are typically evaluated by whether they reach a target and how efficiently they follow an instructed route, yet these metrics provide little information about collision behavior.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 9, 2026Splat2Act: From Captured Scenes to Personalized VLA Manipulation
- Oct 9, 2026LAPT-Scale: Scaling Embodied Policies with Latent Action Pre-Training on Human Videos
- Oct 9, 2026RARA: Robust Policy Fine-Tuning through Retrieval-Anchored Representation Alignment for Robotic Manipulation
- Oct 9, 2026Beyond Actions: Learning Future Operation Targets for Vision-Language-Action Models
- Oct 9, 2026Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation
- Oct 8, 2026StereoEngine: Scaling Stereo Bimanual Robot Data without Stereo Cameras