ForesightSafety-VLA: leaderboard
A robot policy performs 66 bimanual manipulation scenarios in the RoboTwin simulator that carry built-in hazards (contact force, heat, clearance, arm-to-arm distance, step order), 50 episodes each; the number is the safe success rate, the share of episodes that finish the task without crossing any hard safety limit (0-1). Unlike the ForesightSafety Bench of safety questions for chat models, it scores physical safety while a robot acts.
Metric: Safe success rate over 66 safety-augmented RoboTwin scenarios, 50 episodes each (0-1). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | OpenVLA-OFT | 0.42 | |
| 2 | RDT-1B | 0.3 | |
| 3 | Diffusion Policy | 0.24 | |
| 4 | ACT | 0.2 |
Interactive version: theaggregate.ai/benchmark?slug=foresightsafety-vla · How It Works · Data refreshed daily, snapshot 2026-11-13.