Datasets:
metadata
license: mit
task_categories:
- robotics
tags:
- failure-analysis
- pick-and-place
- isaac-sim
- franka-panda
- ur5e
- ur3e
- ur10e
- domain-randomization
- latin-hypercube-sampling
- adaptive-sampling
- physical-ai
size_categories:
- 10K<n<100K
language:
- en
- ko
RoboGate Failure Dictionary
50,000+ Physics-Validated Pick & Place Failure Patterns across 4 Robots (Franka Panda, UR5e, UR3e, UR10e)
A structured database of robot AI failure patterns collected from NVIDIA Isaac Sim physical simulations using Two-Stage Adaptive Sampling. Each experiment records the exact conditions under which a robot succeeded or failed at Pick & Place tasks.
Quick Stats
| Franka Uniform | Franka Boundary | UR5e | UR3e | UR10e | Combined | |
|---|---|---|---|---|---|---|
| Experiments | 10,000 | 10,000 | 10,000 | 10,000 | 10,000 | 50,000+ |
| Success Rate | 33.3% | 63.8% | 74.3% | 10.0% | 0.0% | — |
| Franka Combined | — | 48.6% | — | — | — | — |
| Risk Model AUC | 0.65 | 0.777 | — | — | — | 0.777 |
| Sampling | Uniform LHS | Boundary LHS | Uniform LHS | Uniform LHS | Uniform LHS | Two-Stage |
Key Findings
- friction × mass interaction z = -10.00 — strongest predictor of failure
- Friction threshold: 0.492 ± 0.031 — below this, failure cascades
- Mass > 0.93 kg → Both robots fail at < 40% SR (universal danger zone)
- Boundary equation: μ*(m) = (1.469 + 0.419m) / (3.691 - 1.400m)
- AUC improved 0.65 → 0.777 (+19.5%) with boundary-focused sampling
- Failure mode transition: friction↓ → timeout → collision → grasp_miss
Two-Stage Adaptive Sampling
Stage 1 — Uniform Exploration (40,000)
- Franka Panda 10K + UR5e 10K
- Latin Hypercube Sampling for uniform parameter space coverage
- Identified boundary regions and initial risk model (AUC 0.65)
Stage 2 — Boundary-Focused (10,000)
- Franka Panda only, targeting boundary/transition regions
- Concentrated sampling near friction threshold 0.492
- Revealed failure mode transitions invisible to uniform sampling
- Boosted Risk Model AUC to 0.777 (+19.5%)
Universal Danger Zones (mass > 0.93 kg)
| Mass Range | Franka SR | UR5e SR |
|---|---|---|
| 0.93 – 1.23 kg | 21.4% | 30.9% |
| 1.23 – 1.52 kg | 14.9% | 25.3% |
| 1.52 – 1.82 kg | 12.5% | 28.9% |
| 1.82 – 2.11 kg | 6.6% | 28.1% |
Usage
from datasets import load_dataset
ds = load_dataset("liveplex/robogate-failure-dictionary")
print(ds["train"][0])
# Filter danger zones
danger = ds["train"].filter(lambda x: x["zone"] == "danger")
print(f"Danger zones: {len(danger)}")
Parameter Space
| Parameter | Range | Scale | Paper |
|---|---|---|---|
| friction | 0.05 – 1.2 | log-uniform | SIMPLER 2024 |
| mass | 0.05 – 2.0 kg | log-uniform | SIMPLER 2024 |
| com_offset | 0.0 – 0.40 | uniform | Suction Grasp 2025 |
| size | 0.02 – 0.12 m | uniform | SIMPLER 2024 |
| ik_noise | 0.0 – 0.04 rad | uniform | ICRA Sim2Real 2025 |
| obstacles | 0 – 4 | integer | RoboFAC 2025 |
| shape | 5 types | categorical | Grasp Anything 2024 |
| placement | 14 types | categorical | ALEAS 2025 |
Research Foundations
| Design Choice | Paper | Year |
|---|---|---|
| Two-Stage Adaptive Sampling | ALEAS | 2025 |
| friction × mass interaction | SIMPLER | CoRL 2024 |
| Failure taxonomy | RoboFAC | NeurIPS 2025 |
| Cross-robot validation | RoboMIND | RSS 2025 |
| UR-specific failures | Guardian | ICRA 2025 |
| Confidence intervals | SureSim | Badithela 2025 |
| GPU simulation | Isaac Lab | NVIDIA 2025 |
| Grasp evaluation | Isaac Sim Grasping SDG | NVIDIA 2025 |
VLA Benchmark — 4-Model Leaderboard
Four VLA models evaluated on RoboGate's 68-scenario adversarial suite. All scored 0% SR — including NVIDIA's official GR00T N1.6.
| Model | Params | SR | Confidence | Failure Pattern |
|---|---|---|---|---|
| Scripted Controller | — | 100% (68/68) | 76/100 | — |
| GR00T N1.6 (NVIDIA) | 3B | 0% (0/68) | 1/100 | grasp_miss + collision |
| OpenVLA (Stanford + TRI) | 7B | 0% (0/68) | 27/100 | grasp_miss dominant, 0 collision |
| Octo-Base (UC Berkeley) | 93M | 0% (0/68) | 1/100 | grasp_miss 79%, collision 21% |
| Octo-Small (UC Berkeley) | 27M | 0% (0/68) | 1/100 | grasp_miss 79.4%, collision 20.6% |
Model size is not the bottleneck — even NVIDIA's flagship 3B model cannot bridge the training-deployment distribution gap.
Leaderboard: robogate.io/vla · Paper: arXiv:2603.22126
Citation
@dataset{robogate_failure_dictionary_2026,
title={RoboGate Failure Dictionary: 30K Physics-Validated Pick & Place Failure Patterns},
author={RoboGate Team},
year={2026},
url={https://huggingface.co/datasets/liveplex/robogate-failure-dictionary},
note={Franka Panda + UR5e, Two-Stage Adaptive Sampling, AUC 0.777}
}
Links
- GitHub: liveplex-cpu/robogate-failure-dictionary
- RoboGate Platform: liveplex-cpu/robogate