opentau.policies.cosmos3_nano.modeling_cosmos3_nano

cosmos3_nano: the cosmos3 flow-matching VLA policy on a frozen Qwen3-VL-8B reasoner.

Architecturally identical to policies/cosmos3/modeling_cosmos3.py — the model, prefix/suffix pipeline, flow-matching loss, and sampler are all inherited unchanged. The Cosmos3-Nano reasoner keeps the exact KV interface the expert cross-attention consumes (8 KV heads x head_dim 128), so the only difference is the backbone checkpoint (TensorAuto/cosmos3-reason-8b) and depth (36 layers), both carried by Cosmos3NanoConfig.

Classes

Cosmos3NanoPolicy(config[, ...])

OpenTau wrapper for cosmos3_nano — Cosmos3Policy with the Nano backbone defaults.

class opentau.policies.cosmos3_nano.modeling_cosmos3_nano.Cosmos3NanoPolicy(config: Cosmos3Config, per_dataset_stats: list[dict[str, dict[str, Tensor]]] | None = None, dataset_names: list[str] | None = None, qwen3vl_config: Qwen3VLConfig | None = None)[source]

Bases: Cosmos3Policy

OpenTau wrapper for cosmos3_nano — Cosmos3Policy with the Nano backbone defaults.

config_class: alias of Cosmos3NanoConfig

name: None = 'cosmos3_nano': The name of the policy. Must be defined in subclasses.