Skip to content

Policy adapters

Policies implement tower_eval.PolicyAdapter:

def reset(*, seed=None, metadata=None): ...
def act(observation, info=None): ...
def close(): ...

act may return one action or tower_eval.ActionChunk(actions, dt_s, metadata). The evaluator executes the supplied chunk up to the episode limit and checks its timestep. TOWER owns this lightweight type; no serving package is required.

LocalPolicyAdapter wraps callable/act policies and XPolicyLab-style models with update_obs, get_action and argument-free reset. It uses the same encoder observation mapping, action decoder and exact anchor validation as the WebSocket client. Supply all camera/state fields; raw CPU qpos is insufficient. Model-style joint actions use standard XPolicyLab dictionaries and are decoded into joint16 encoder endpoints; specify action_mode="eef_delta" for tagged TOWER encoder EEF targets. prepare_case receives the episode seed when the model supports it. See the random CPU model.

Remote models run through XPolicyLabAdapter:

[policy]
adapter = "xpolicylab"
server = "ws://localhost:19000"
timeout_s = 120

The TOWER ↔ XPolicyLab contract defines required modalities, frames, action modes, batching and deployment. TOWER contains no policy server. Use one runtime process per server. The Isaac runtime uses atomic INFER followed by optional get_prediction(env_idx) calls per environment. Expose get_identity for startup authorization and retain identity on every endpoint and full prediction. See the reference for error categories and lifecycle.

See the policy interface for units, frames and lifecycle.