"""Unrehearsed observation-conditioned nonparametric action retrieval.""" from pathlib import Path import numpy as np import cv2 _mem=None _phase=0. def reset_episode(): global _phase _phase=0. def act(obs): global _mem,_phase if int(obs.get('step_id',1))>=500:return None if _mem is None: with np.load(Path(__file__).with_name('memory.npz')) as z:_mem={k:z[k] for k in z.files} try: q=np.asarray(obs['proprioception']['agent/qpos']).reshape(-1) v=np.asarray(obs['proprioception']['agent/qvel']).reshape(-1) fs=[] for cam in ['front_camera','left_camera','top_camera']: im=np.asarray(obs['rgb'][cam]);im=im[0] if im.ndim==4 else im fs.append(cv2.resize(im,(16,12),interpolation=cv2.INTER_AREA).reshape(-1)/255.) f=np.concatenate(fs) if q.shape!=(23,) or v.shape!=(23,) or not np.isfinite(np.r_[q,v,f]).all():return np.zeros(22,dtype=np.float32) except (KeyError,ValueError,cv2.error):return np.zeros(22,dtype=np.float32) # Local phase continuity limits ambiguous revisits; current RGB and joints choose every action. ix=np.flatnonzero((_mem['phase']>=max(0,_phase-.035))&(_mem['phase']<=_phase+.09)) dist=np.mean(((_mem['q'][ix]-q)/.35)**2,axis=1)+.04*np.mean(((_mem['v'][ix]-v)/2)**2,axis=1)+12*np.mean((_mem['image'][ix]-f)**2,axis=1) j=ix[int(np.argmin(dist))] # No timed advancement and no demo/seed commitment. Stall remains possible. _phase=min(.998,max(_phase+1/440.,float(_mem['phase'][j]))) a=_mem['a'][j].copy() return np.clip(a,-1,1).astype(np.float32)