SimToReal Agent

Problem: Re-tasking a factory robot is slow and risky. AI can write a robot program in seconds, but a program that looks valid can still drop a part, hit the table or reach for something out of range. On a real line that means scrap, damaged tooling and downtime, so nobody should run an untested AI program on a real machine. User: the cell operator or automation engineer who must sign off that a new robot program is safe. How it works: The operator types a task such as "Sort both parts by color". An LLM on AMD Token Factory turns it into explicit goals and writes eight candidate pick-and-place programs, each checked against the allowed actions and objects. All eight run in parallel in a MuJoCo simulation of a Franka Panda arm, which records video, a final camera image, collisions and dropped parts. A safety-first scorer rejects any program that collides, drops a part or misses the goal, and ranks the rest. A vision model on AMD checks the best program's final image; if it disagrees with the simulator or cannot tell, the plan is flagged for human review. The operator reviews the evidence and approves, and only then is the program exported. The app never commands a real robot. AMD: All inference runs on AMD Token Factory: DeepSeek-V4-Flash plans, DeepSeek-V4.1-Flash reviews images. An AMD panel shows the models and measured times for every run. Results: In live runs, 24 AI-written programs for three tasks were tested. 12 passed and 12 unsafe or impossible ones were caught in simulation, each with a readable reason. For an out-of-reach part, all eight were rejected and no plan was offered. Planning eight programs took 6 to 8 seconds.