Cosmos 3 Edge puts a 4B world model on Jetson Thor. The reported trade-off: 22.9% RoboLab success, compared with Nano’s 36.8%, at a fraction of the inference compute.
Why it matters: For teams weighing on-robot deployment against benchmark success, the article makes the trade-off concrete: Edge fits and runs on Jetson Thor, while its reported RoboLab success is lower than Nano’s. It also describes simulation as a way to evaluate behavior before a physical-robot test.
SPADE trains an environment designer and a reasoning agent, using a hint-based regret signal to target feasible tasks near the agent’s capabilities. The captured v3 abstract is a later revision from 2026-08-31.
Why it matters: Training environments can evolve with a learning agent.
RISE estimates the benefit of continuing a simulated rollout and weighs it against compute; the authors also describe a verified counterfactual driving dataset.
Why it matters: Adaptive rollout budgets may focus planning on difficult driving scenes.