Running a robot policy on-device introduces two constraints: memory and control latency.
Our new technical blog shows how to post-train the 4B Cosmos 3 Edge model using Cosmos3-DROID, serve it on Jetson Thor, run receding-horizon control, and evaluate it in RoboLab. On Jetson AGX Thor T5000 at 640×360 and 15 Hz, each action chunk is generated in about 1.53 seconds and covers roughly 2.13 seconds of robot motion.
For developers working beyond robot control, the August 20 livestream covers VLM post-training for video reasoning and Cosmos 3 Super step distillation for faster synthetic data and world generation, with examples from Aigen and Linker Vision.
Choose how you want to get started:
📖 Read the Technical Blog
🤗 Explore the Cosmos 3 Model Collection
🧑💻 Get Models and Code on GitHub
📺 Tune in to the Cosmos Labs Livestream on August 20
Building a policy with your own robot data? Share your embodiment, dataset, VLA baseline, results, or current blocker with the community in the Cosmos Discord channel.