Sinewgrid Autonomous trains world models for robots and autonomous systems.

Robots and vehicles fail on the rare cases: an unusual obstacle, odd lighting, a sensor fault, a situation that happens once in a million miles. Logged runs rarely contain enough of them to train on, and collecting them for real is slow and unsafe. A world model learns to predict how a scene evolves from video and three-dimensional sensor data, and once trained it can generate new scenarios, including hard ones, for a policy to practice on. Sinewgrid Autonomous would build these models. The compute demand is large on both sides. Training on video moves far more data than training on text, so storage and interconnect matter, and generating a long, consistent scene takes many iterative steps for every clip.

The work is a loop that does not stop. Logs the partner is entitled to use are ingested and auto-labeled, with a sample checked by people, the world model is trained, rare and hard scenarios are generated and checked for realism, and policies trained with them are scored on real held-out cases. We keep a hard line between generated and real: generated scenes are labeled as such, and no policy is judged on generated data alone. We would report a realism check, the change in a policy’s score on real held-out cases after training with generated scenarios, and the cost per generated hour. Training runs on a pool with fast interconnect, generation on interruptible capacity and evaluation on its own pool. This domain is at the exploring stage.