dp / code /hb_baselines_v1 /baselines /adapters.py
BIue-Jay's picture
Add B6 DP batch-512 step-4000 checkpoint with resume state (2026-11-17)
ab621ab verified
Raw History Blame Contribute Delete
2.57 kB
"""The G1 evaluation adapters of the baselines (evaluation/README.md#the-model-interface): `make_model('fm_dit')`
returns the `Model` class that `evaluation/policy/FMDiT/model.py` exports (likewise DP, ACT).
Observation: the current ego_view and chest frames (480x640 RGB, resized as the training cache) and state44; no past
actions or earlier frames (the baselines read one frame). Output: the checkpoint's whole chunk (16 rows of action76,
row 0 executed at the request's frame), `execution_steps` rows of it executed before the next query (deploy.yml
`config: execution_steps`, default the whole chunk), the request's seed for the sampling noise (FM-DiT, DP).
"""
from __future__ import annotations
from .runtime import BaselinePolicy
WIDTH = {'sonic_dfx6': 76}
def require(condition, message):
if not condition:
raise RuntimeError(message)
def make_model(policy_name: str):
class Model:
history_length = 1 # the current head and chest frames only
video_frames = 0
def __init__(self, cfg):
self.policy = BaselinePolicy(cfg['checkpoint_path'], device=cfg.get('device', 'cuda:0'),
policy=policy_name, robot='g1')
contract = self.policy.contract
require(cfg['action_space'] in WIDTH and contract['action_space'] == cfg['action_space']
and self.policy.action_dim == WIDTH[cfg['action_space']] and self.policy.state_dim == 44
and list(self.policy.cameras) == ['ego_view', 'chest'],
f"checkpoint contract {contract['name']} does not serve {cfg['action_space']}")
extra = cfg.get('config') or {}
if extra.get('check_task', True):
require(cfg['task'].strip() == (self.policy.task_text or '').strip(),
f'task text {cfg["task"]!r} is not the training task {self.policy.task_text!r}')
self.execution_steps = int(extra.get('execution_steps') or self.policy.horizon)
require(1 <= self.execution_steps <= self.policy.horizon, f'execution_steps {self.execution_steps}')
self.info = dict(self.policy.info(), execution_steps=self.execution_steps)
def get_action(self, obs):
action, _ = self.policy.predict(obs['images'], obs['state'], seed=obs['seed'])
return dict(action=action, execution_steps=self.execution_steps,
sampling_steps=self.policy.sampling_steps)
Model.__name__ = Model.__qualname__ = f'Model_{policy_name}'
return Model