Astra teaches a robot arm to screw in a lightbulb in Isaac Sim
Reproducing the EmbodiedSWE benchmark, GPT-6 Astra wrote a controller in 4 hours for about $85–120 in API costs, while Claude Opus 5.5 failed to seat the thread in both of its setups.
We had GPT-6 Astra, an AI coding agent, teach a robot arm to screw in a lightbulb, entirely in NVIDIA Isaac Sim. The task: we reproduced the new EmbodiedSWE benchmark on our own rented GPUs. A simulated Franka arm has to pick up a bulb lying on its side, stand it upright, set it on a socket, and screw it down along a physically simulated thread until it seats. The agent gets a simulator, a camera, and 4 hours to write the controller from scratch. By the numbers: - 4 hours - 7 work sessions - 289 commands - 22 checkpoints - 47.7M input tokens (43.8M cached) - 42.7K output tokens - ~$85–120 in API calls at list price - ~$4 of GPU time (5.2 hours on a single RTX 4090) The model is ~95% of the cost. It read ~1,100 tokens for every token it wrote. For comparison, Claude Opus 5.5 on the same task and setup got the bulb upright on the socket but never got the thread to catch: 0 of 2 setups seated so far. Next: turning Astra's working solution into hundreds of verified robot demonstrations for training robot models, all generated in simulation.