Linslab's picture
Move frozen Astra policies to astra/<task>/ (DualArmHammerHandover-v1 .. DualArmPushBox-v1)
67e0b67 verified
Raw History Blame Contribute Delete
1.38 kB
"""Fixed image descriptors; no neural model or trained weights."""
import cv2
import numpy as np
CAMERAS=('top_camera','front_camera')
PATHS=tuple('agent/panda-%d/%s'%(r,p) for r in range(2) for p in ('qpos','qvel'))
def image_features(rgb):
out=[]
for name in CAMERAS:
a=np.asarray(rgb[name])
if a.ndim==4 and a.shape[0]==1:a=a[0]
if a.ndim!=3 or a.shape[2]!=3 or a.dtype!=np.uint8:raise ValueError('RGB must be uint8 HWC or 1HWC')
h,w=a.shape[:2]
# Central workspace, retaining robot fingers and tile positions.
a=a[int(.22*h):int(.83*h),int(.13*w):int(.87*w)]
small=cv2.resize(a,(32,32),interpolation=cv2.INTER_AREA).astype(np.float32)/255.
green=np.maximum(small[:,:,1]-(small[:,:,0]+small[:,:,2])/2,0)
white=np.maximum(small.min(axis=2)-.55,0)
out.extend([small.ravel(),(3*green).ravel(),(2*white).ravel()])
return np.concatenate(out)
def proprio_features(p):
parts=[]
for path in PATHS:
a=np.asarray(p[path],dtype=np.float32).reshape(-1)
if a.size!=9 or not np.isfinite(a).all():raise ValueError('invalid proprioception')
# Physical scales chosen manually, not optimized.
scale=np.array([.3]*7+[.02,.02],np.float32) if path.endswith('qpos') else np.array([2.]*9,np.float32)
parts.append(a/scale)
return np.concatenate(parts)