RoboTrustBench - Constraint-Sensitive: leaderboard
RoboTrustBench (2026) tests whether image-to-video world models generate trustworthy robot-arm manipulation videos from a real DROID starting frame and a language instruction. This board is the human-rated overall on feasible but constrained instructions (ambiguous references, occlusion, distractors, obstacles and trajectory constraints) from a stratified subset of 10 pairs per fine-grained subcategory: three raters score 12 criteria from 1 to 5, averaged and normalized to 0-100; higher is better.
Metric: Overall average (0-100) of the five dimension means over 12 human-rated criteria on the Constraint-Sensitive (feasible but ambiguous, occluded, cluttered or trajectory-constrained instruction) scenario, stratified human-evaluation subset (10 instruction-image pairs from each of the 18 fine-grained subcategories, 180 in all; three raters score each generated video 1 to 5 per criterion, normalized to 0-1 and printed x100 here): an image-to-video world model generates a robot-arm manipulation video from a real DROID initial frame and instruction, rated for scene entity alignment, spatiotemporal consistency, interaction rationality, task execution and visual quality; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kling 2.6 | 75.9 | |
| 2 | Veo 3.1 Fast | 65.7 | |
| 3 | Cosmos-Predict2.5-14B | 61.2 | |
| 4 | Wan2.2-I2V-A14B | 58.6 | |
| 5 | LingBot-World | 57.8 | |
| 6 | Cosmos-Predict2.5-2B | 57 | |
| 7 | HunyuanVideo-1.5 | 50.8 |
Interactive version: theaggregate.ai/benchmark?slug=robotrustbench-constraint-sensitive · How It Works · Data refreshed daily, snapshot 2026-11-13.