opentau.policies.cosmos3

cosmos3: a flow-matching Vision-Language-Action policy.

cosmos3 pairs a frozen Qwen3-VL-32B backbone – the reasoning tower of NVIDIA Cosmos3-Super, extracted into a standalone Qwen3-VL-32B checkpoint – with a custom, sub-1B Qwen3-style flow-matching action expert in the π0.5 dual-stream style: the backbone encodes images + the language prompt once (prefix), and the trainable expert cross-attends to the backbone’s per-layer key/value cache to denoise a continuous action chunk. Continuous actions only (MSE flow matching) – no FAST discrete-action tokens and no subtask/response head.

Modules

configuration_cosmos3	Configuration for the cosmos3 policy.
modeling_cosmos3	cosmos3: a Vision-Language-Action flow-matching policy on a frozen Qwen3-VL-32B reasoner.
qwen3vl_with_expert	Qwen3-VL backbone + custom flow-matching action expert (cosmos3).