Benchmarks/RoboLab-120

RoboLab-120

SimulationCode checkReleased Apr 2026Index weight 0.95
NVIDIA Seattle Robotics LabMaintainer ↗

120 tasks. Difficulty (simple, moderate, complex) and skill (visual, relational, procedural), plus reworded instructions.

As of November 17, 2026, AMP1 leads with 86.0, followed by SenseNova-RoboRSI (57.1) and Lakeside (51.0). Official leaderboard ↗ · Paper ↗

Leaderboard

Tasks
120
Attempts / task
10
Median cost / task
$0.068
Median time / task
72 s

Overall vs speed

Published time per task attempt

Faster, higher scoring10.020.030.015 s20 s30 sTime per task (faster ←)Overall (better ↑)π0.5

Cost isn’t published

RoboLab-120 doesn’t report cost per attempt for enough models to chart. The table shows Playground’s estimates, marked est.; Methodology explains how they’re made.

Success rate (%), 10 attempts per task.

#Model
1AMP1Not counted in the index86.086.887.779.1——
2SenseNova-RoboRSI57.159.456.250.6$17.74est.65 sest.
3LakesideNot counted in the index51.057.644.142.4——
4BiMind v0.5Not counted in the index49.251.652.632.4——
5Praxis-146.246.947.940.0$0.064est.70 sest.6FLUX 3 Action42.949.139.228.2$0.065est.72 sest.7Cosmic42.346.640.031.8$0.066est.72 sest.
8HiDream-O1-EmbodiedNot counted in the index39.943.137.732.9——
9Atomic-WAMNot counted in the index39.644.440.519.4——
10OASIS WAMNot counted in the index39.041.437.932.4——
11Cosmos3-Nano-Policy36.840.635.425.3$0.068est.74 sest.12Phoenix34.442.825.622.9$0.069est.76 sest.
13BiMind v0.1Not counted in the index33.338.832.814.1——
14VoLoNot counted in the index28.232.725.617.6——
15π0.528.029.731.513.5$0.015est.17 s16DreamZero25.726.130.014.1$0.073est.80 sest.17Cosmos3-Edge-Policy22.925.623.311.8$0.074est.81 sest.18π0-FAST15.520.213.32.9$0.018est.19 s19GR00T N1.67.28.87.90.00$0.080est.88 sest.20π05.07.23.60.00$0.029est.32 s21PaliGemma-binning3.43.44.90.00$0.082est.90 sest.

Tasks

120 tasks. Hover a task to play its clip (15 of 120).

120 tasks · 15 with clips

Methodology

Scoring
Code check
Time cap / attempt
1 min 31 s
Control rate
15 Hz
Models in index
13 of 21

Scoring. Simulator end-state check per task, plus a subtask progress score.

Who runs it. NVIDIA runs or cross-checks entries.

Robots. Simulated Franka (DROID setup).

Time per attempt. Published or measured for 3 of 13 models. Estimated otherwise: time cap 91 s × (1 − success/2).

Cost per attempt. Not published. Estimated otherwise: one H100 at $3.29 per hour (Lambda on-demand, Oct 2026) × time; GPT-6 Astra's median published cost per attempt on other boards. GPU price source ↗