Automated validation pipeline powered by NVIDIA Isaac Sim
Why we chose these parameters — each backed by peer-reviewed research
| Parameter | Range | Scale | Paper | Why |
|---|---|---|---|---|
| friction | 0.05 – 1.2 | log-uniform | SIMPLER 2024 | Friction-success correlation r=+0.36 (#1) |
| mass | 0.05 – 2.0 kg | log-uniform | SIMPLER 2024 | Reflects Franka payload limits |
| com_offset | 0.0 – 0.4 | uniform | Suction Grasp 2025 | CoM bias → grasp stability shift |
| size | 0.02 – 0.12 m | uniform | SIMPLER 2024 | Failure rate spikes below 4cm |
| ik_noise | 0.0 – 0.04 rad | uniform | ICRA Sim2Real 2025 | Sim-to-Real control error simulation |
| obstacles | 0 – 4 | integer | RoboFAC 2025 | Collision rate doubles at 3+ obstacles |
| shape | 5 types | categorical | Grasp Anything 2024 | box, cylinder, sphere, L, irregular |
| placement | 14 types | categorical | ALEAS 2025 | rotation / tilt / edge placements |
Two-Stage Adaptive Sampling: Stage 1 uniform LHS 20K + Stage 2 boundary-focused LHS 10K → AUC 0.65 → 0.777 (+19.5%)
6 failure types based on RoboFAC (2025)
Gripper approached but failed to grasp the object
size < 3cm, approach_angle > 60°
Robot collided with table, object, or obstacle
obstacles ≥ 3, cluttered scene
Object dropped during transport after grasp
friction < 0.2, mass > 1.5kg
Task not completed within time limit
complex IK solution, non-vertical approach
Slippage detected during grasp (object not dropped)
friction < 0.3, com_offset > 0.2
Object not placed accurately at target position
reach_ratio > 0.82, IK noise
Same experiments on two robots → discover universal danger zones
Franka Panda (7DOF)
20,000
experiments · SR 48.6% (uniform 10K + boundary 10K)
danger zone: 7,808
UR5e (6DOF)
10,000
experiments · 74.3% success rate
UR5e PickPlaceController + SurfaceGripper (suction)
danger zone: 2,570
UR3e (6DOF)
10,000
Lightweight robot validation
NEW
UR10e (6DOF)
10,000
Heavy-duty robot validation
NEW
Universal danger zone: mass > 0.93 kg → both robots SR < 40%. Boundary equation: μ*(m) = (1.469 + 0.419m) / (3.691 - 1.400m)
Wilson Score Interval — finite sample confidence bounds (SureSim 2025)
50,000+
Samples
±0.6%
95% CI margin
p < 0.001
Statistical significance
Wilson Score: p̂ ± z·√(p̂(1-p̂)/n + z²/4n²) / (1 + z²/n)
n=50,000, p̂=0.557, z=1.96 → CI = [0.553, 0.561]
Papers that RoboGate is built on
friction × mass joint sampling, Sim-to-Real gap quantification (24-30%)
Latin Hypercube Sampling — 2-3× space coverage vs. random
IK noise injection for domain randomization
Wilson Score Interval — finite sample confidence bounds
6-type failure classification taxonomy
Center of Mass offset effect on grasp stability
5 object shapes (box, cylinder, sphere, L, irregular)
GPU parallel environments (4096 envs), Newton Physics engine