"""Fixed, non-neural descriptors of legal observations.""" import cv2 import numpy as np CAMERAS = ('top_camera', 'front_camera') PATHS = tuple(f'agent/panda-{i}/{s}' for i in range(2) for s in ('qpos','qvel')) def image_feature(value): a = np.asarray(value) if a.ndim == 4 and a.shape[0] == 1: a = a[0] if a.ndim != 3 or a.shape[2] != 3: raise ValueError('Expected RGB HWC or 1HWC') if not np.isfinite(a).all(): raise ValueError('Nonfinite RGB') small = cv2.resize(a.astype(np.float32)/255., (24,32), interpolation=cv2.INTER_AREA) # Cyan boundary pixels are thin; pooling preserves their location. x=a.astype(np.float32)/255. cyan=((x[:,:,1]-x[:,:,0]>.10)&(x[:,:,2]-x[:,:,0]>.10)).astype(np.float32) cyan=cv2.resize(cyan,(24,32),interpolation=cv2.INTER_AREA) return np.concatenate([small.reshape(-1), 4*cyan.reshape(-1)]).astype(np.float32) def proprio_feature(obs): out=[] for p in PATHS: x=np.asarray(obs['proprioception'][p],dtype=np.float32).reshape(-1) if x.size!=9 or not np.isfinite(x).all(): raise ValueError('Invalid proprioception') scale=np.array([.35]*7+[.025]*2) if p.endswith('qpos') else np.array([2.]*7+[.2]*2) out.extend(x/scale) return np.asarray(out,dtype=np.float32)