Datasets:
🚀 50K+ Failure Episodes, 4 Robots, 6 VLA Models — All Scored 0%
#2
by liveplex - opened
Major Update: 50K+ Failure Episodes Across 4 Robots
The RoboGate Failure Dictionary has been expanded to 50,000+ episodes across 4 industrial robot arms:
| Robot | Episodes | Success Rate |
|---|---|---|
| Franka Panda (uniform) | 10,000 | 33.3% |
| Franka Panda (boundary) | 10,000 | 63.8% |
| UR5e | 10,000 | 74.3% |
| UR3e | 10,000 | 10.0% |
| UR10e | 10,000 | 0.0% |
6-Model VLA Leaderboard — All 0%
We tested 5 state-of-the-art VLA models on our 68 adversarial pick-and-place scenarios using real NVIDIA Isaac Sim physics:
| Model | Params | Success Rate |
|---|---|---|
| Scripted (IK baseline) | — | 100% |
| GR00T N1.6 (NVIDIA) | 3B | 0% |
| SmolVLA Base (HuggingFace) | 450M | 0% |
| OpenVLA (Stanford+TRI) | 7B | 0% |
| Octo-Base (UC Berkeley) | 93M | 0% |
| Octo-Small (UC Berkeley) | 27M | 0% |
Key finding: Scaling from 27M to 7B parameters (260x) made zero difference. The bottleneck is training data–deployment environment mismatch, not model size.
Resources
- Paper: arXiv 2603.22126
- Leaderboard: robogate-vla-leaderboard
- Code: github.com/liveplex-cpu/robogate
- Website: robogate.io
We welcome community contributions — submit your VLA model results to the leaderboard!