Overview

A console that monitors a parallel simulation farm, synthetic-data generation, and the sim-to-real transfer gate on one screen.

Project Background & Purpose

The customer was bottlenecked on hardware cell time, so policy candidates had to be screened in simulation first. Sim and hardware results were tracked separately, decisions on which policy to run on hardware came late, and farm utilization was invisible. The project runs 64 Isaac Lab environments through a Ray job queue and puts domain-randomization parameters, gate pass rates, and hardware run-off results on the same console. Simulation numbers are used only as a screening criterion; acceptance is decided from hardware run-off records.

Components

• 64 Isaac Lab parallel environments (RTX PRO 6000 ×3 render/physics workers, H200 training node)
• Ray job queue, checkpoints, domain randomization (lighting 300–1,500 lux, friction 0.45–0.95, mass 4.8–7.4 kg)
• Transfer gate and hardware-cell controller-in-the-loop (PLEM to ROS 2 bridge, W-RC)

Process Steps

1. 1Pretrain the pick-and-place policy in batch #217 toward 50,000 episodes (38,912 so far, 87.4 % success)
2. 2Register only checkpoints above 90 % sim success as hardware run-off candidates
3. 3Pass if 64 hardware run-offs land within 5 %p of sim; otherwise return to the rework queue