SleepWalk - Trajectory Efficiency: leaderboard

SleepWalk (2026) tests instruction-guided vision-language navigation: from views of a text-to-3D reconstructed scene and an easy, medium or hard instruction, a model predicts a 3D waypoint trajectory that a GPT-5-mini judge rates from 1 to 5. This board scores whether the path is reasonably direct, as the rating divided by 5 (0.2 to 1) averaged over the three tiers; higher is better.

Metric: Trajectory efficiency (whether the route is reasonably direct without unnecessary detours): GPT-5-mini judge rating on 1 to 5 divided by 5 (0.2 to 1), averaged over valid trajectories within each difficulty tier and then over the easy, medium and hard tiers with equal weight; each model predicts a 3D waypoint trajectory from a top-down and an oblique view of a Hunyuan3D-reconstructed scene and a navigation instruction, zero-shot; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5 Mini0.64#364
2Gemini Robotics-ER 1.50.58#291
3Qwen3-VL (SleepWalk checkpoint unspecified)0.47

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=sleepwalk-trajectory-efficiency · How It Works · Data refreshed daily, snapshot 2026-11-21.