ForesightSafety-VLA - Safety Cost: leaderboard
The same 66 RoboTwin hazard scenarios; the number is the cumulative safety cost, the per-step penalty for operating inside the warning margin of a safety limit, averaged over episodes; lower is better. It measures how close to danger the policy works, where the headline only counts tasks finished without a hard violation.
Metric: Cumulative safety cost per step over 66 safety-augmented RoboTwin scenarios (lower is better). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | OpenVLA-OFT | 0.18 | |
| 2 | RDT-1B | 0.29 | |
| 3 | Diffusion Policy | 0.34 | |
| 4 | ACT | 0.39 |
Interactive version: theaggregate.ai/benchmark?slug=foresightsafety-vla-safety-cost · How It Works · Data refreshed daily, snapshot 2026-11-13.