Give your agent SimForge

Configure a sensor rig

Six 120° cameras and a 64-channel roof lidar, validated with 360° camera coverage; a four-second Easterbrook school-crosswalk render at 17:00, 80 frames at 20 fps, contact and parity passing, and one lidar-coverage clip.

Scenario generation

Built on the school frontage and checked in 2D: take 3 hit the pedestrian at 6.36 s, so the agent swept the start times until the lead brakes from 40 km/h to 1.89 m/s (a second pedestrian makes it brake, since authoring has no brake operation yet) and ours steps out from between the parked cars at 3.4 s; the ego slows to 1.62 m/s and lets them cross. Then rendered once in 3D.

Stress-test a policy closed loop

20 of 20 variations fail: 6 crashes, 7 contacts and 7 near-misses. The policy never yields. The worst run, v14, is the actual closed-loop failure, replayed in 3D from a fixed camera over the junction.

Train a policy with PPO

129,024 steps in 207 s on CPU, one training seed (evaluated on seeds 1001 and 1002); held-out return -2.44 → 1.90 (peak 2.93), under the scripted baseline's 2.04. Rendered in 3D: the policy slowing for a cyclist the scripted driver hits.

Synthetic data

4 scenarios × morning, noon and night = 12 clips of 3 s at 20 fps: 720 labelled frames, front-camera RGB, instance and semantic passes, and all 720 roof-lidar scans retained losslessly; all render gates pass.

Set up your agent

Install the CLI, sign in and add the SimForge skills to Claude Code:

$curl -fsSL https://simforge.ai/install.sh | sh$simforge login$simforge skills install --claude

For Codex, end with simforge skills install --codex. The skills, the zip download and other agents are in the agent skills pack docs.