File size: 9,032 Bytes
972f606 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 | """WAM-ready variable-DOF, synchronized multi-arm transitions from accepted rows."""
from collections import OrderedDict
import json
from pathlib import Path
import pickle
import cv2
cv2.setNumThreads(1)
import lmdb
import numpy as np
from scipy.spatial.transform import Rotation
import io
class _ArrayUnpickler(pickle.Unpickler):
# Only NumPy's data constructors used by the existing array serialization.
# No arbitrary modules, callables or application classes are accepted.
def find_class(self, module, name):
allowed = {
('numpy', 'ndarray'), ('numpy', 'dtype'),
('numpy.core.multiarray', '_reconstruct'), ('numpy.core.multiarray', 'scalar'),
('numpy._core.multiarray', '_reconstruct'), ('numpy._core.multiarray', 'scalar'),
('numpy.core.numeric', '_frombuffer'), ('numpy._core.numeric', '_frombuffer'),
}
if (module, name) not in allowed:
raise pickle.UnpicklingError('unsupported serialized class: '+module+'.'+name)
return super().find_class(module, name)
def _loads(data):
return _ArrayUnpickler(io.BytesIO(data)).load()
class TrainingDataset:
def __init__(self,index,max_open=3):
index = Path(index).expanduser().resolve()
self.root = index.parent
if (self.root/'PUBLICATION_REVIEW_IN_PROGRESS.json').exists():
raise RuntimeError('The portable release has not finished technical verification.')
if not isinstance(max_open, int) or isinstance(max_open, bool) or max_open < 1:
raise ValueError('max_open must be a positive integer')
for parent in Path(index).resolve().parents:
if (parent/'TRAINING_REVIEW_IN_PROGRESS.json').exists():
raise RuntimeError('Dataset review is incomplete; use the finalized refinement train.jsonl only.')
self.rows=[json.loads(line) for line in Path(index).read_text().splitlines() if line.strip()]
assert self.rows and all(row.get('accepted_for_training') for row in self.rows),'use an accepted train.jsonl index'
for row in self.rows:
for field in ('render', 'trajectory'):
relative = Path(row[field])
if relative.is_absolute() or '..' in relative.parts:
raise ValueError('release payload paths must be relative to train.jsonl')
resolved = (self.root/relative).resolve()
if self.root not in resolved.parents:
raise ValueError('payload path escapes the release directory')
row[field] = str(resolved)
self.lengths=[row['num_steps'] for row in self.rows];self.ends=np.cumsum(self.lengths)
self.cache=OrderedDict();self.max_open=max_open
def __len__(self):return int(self.ends[-1])
def _open(self,episode):
if episode in self.cache:
entry=self.cache.pop(episode);self.cache[episode]=entry;return entry
row=self.rows[episode];folder=Path(row['render']);meta=_loads((folder/'meta_info.pkl').read_bytes())
assert meta['accepted_for_training'] and meta['views']==row['views']
db=lmdb.open(str(folder/'lmdb'),readonly=True,lock=False,readahead=False);arrays={}
with db.begin() as tx:
for robot in meta['robots']:
name=robot['name']
arrays[name]=dict(state=_loads(tx.get(f'observation/{name}/qpos'.encode())),
velocity=_loads(tx.get(f'observation/{name}/qvel'.encode())),
feedforward_effort=_loads(tx.get(f'action/{name}/joint_feedforward_effort'.encode())),
action=_loads(tx.get(f'action/{name}/joint_position'.encode())),
next_state=_loads(tx.get(f'state/{name}/next_qpos'.encode())))
with np.load(folder/'camera-calibration.npz',allow_pickle=False) as calibration:
camera_poses={name:calibration[name].copy() for name in row['views']}
with np.load(Path(row['trajectory'])/'trajectory.npz',allow_pickle=False) as physical:
for name in row['views']:
mount=meta['camera_config'][name]['mount']
if mount['type']=='world':terminal_pose=np.asarray(mount['world_pose'])
else:
link=physical[f'arm_{mount["arm"]+1}_link_transforms'][-1,mount['link_index']]
rotation=Rotation.from_quat(link[3:])
local=Rotation.from_quat(np.asarray(mount['local_orientation_wxyz'])[[1,2,3,0]])
terminal_pose=np.r_[link[:3]+rotation.apply(mount['local_position']),(rotation*local).as_quat()[[3,0,1,2]]]
camera_poses[name]=np.vstack([camera_poses[name],terminal_pose])
entry=(row,meta,db,arrays,camera_poses);self.cache[episode]=entry
if len(self.cache)>self.max_open:self.cache.popitem(last=False)[1][2].close()
return entry
def __getitem__(self,index):
if not 0<=index<len(self):raise IndexError(index)
episode=int(np.searchsorted(self.ends,index,side='right'))
frame=index-(int(self.ends[episode-1]) if episode else 0)
return self.step(episode,frame)
def step(self,episode,frame,action_horizon=1):
if not 0 <= episode < len(self.rows):raise IndexError(episode)
row,meta,db,arrays,camera_poses=self._open(episode);n=row['num_steps']
if not 0<=frame<n:raise IndexError(frame)
if not isinstance(action_horizon,int) or isinstance(action_horizon,bool) or action_horizon<1:
raise ValueError('action_horizon must be a positive integer')
def image(tx,name,k):
if k==n:rgb=cv2.imread(str(Path(row['render'])/f'terminal_{name}.png'))
else:
key=meta['keys'][f'observation/{name}/color_image'][k];payload=tx.get(key)
if payload is None:raise ValueError('missing RGB image')
rgb=cv2.imdecode(np.asarray(_loads(payload),dtype=np.uint8),cv2.IMREAD_COLOR)
if rgb is None or rgb.shape!=(480,640,3):raise ValueError('invalid RGB image')
return rgb[:,:,::-1].copy()
with db.begin() as tx:
images={name:image(tx,name,frame) for name in row['views']}
next_images={name:image(tx,name,frame+1) for name in row['views']}
indices=np.minimum(np.arange(frame,frame+action_horizon),n-1)
robots={}
for description in meta['robots']:
name=description['name'];a=arrays[name]
units=description.get('joint_units')
if units is None:
# Early accepted pilots predate the explicit unit field. Their
# unchanged native models have revolute arms and two prismatic
# fingers; all 13 Robotiq mechanism joints are revolute.
count=len(description['joint_names'])
units=['rad']*count if description['kind']=='robotiq' else ['rad']*(count-2)+['m','m']
robots[name]=dict(kind=description['kind'],joint_names=description['joint_names'],
joint_units=units,
actuator_effort_limits=description.get('actuator_effort_limits'),
drive_stiffness=description.get('drive_stiffness'),drive_damping=description.get('drive_damping'),
actuated_mask=np.asarray(description.get('actuated_mask',[True]*len(description['joint_names'])),dtype=bool),
velocity=a['velocity'][frame].astype(np.float32),feedforward_effort=a['feedforward_effort'][frame].astype(np.float32),
state=a['state'][frame].astype(np.float32),action=a['action'][frame].astype(np.float32),
next_state=a['next_state'][frame].astype(np.float32),action_chunk=a['action'][indices].astype(np.float32))
return dict(episode=row['episode'],task=row['task'],instruction=row['instruction'],frame=frame,timestamp_s=frame/30,
robots=robots,images=images,next_images=next_images,terminal=frame==n-1,
cameras={name:dict(intrinsics=np.asarray(meta['camera_config'][name]['intrinsics'],dtype=np.float32),
pose_world_wxyz=camera_poses[name][frame].copy(),next_pose_world_wxyz=camera_poses[name][frame+1].copy(),
axes='ros',resolution=[640,480]) for name in row['views']},
action_chunk_mask=np.arange(frame,frame+action_horizon)<n,
action_semantics='absolute native joint position commands; radians and metres as named')
def close(self):
for entry in self.cache.values():entry[2].close()
self.cache.clear()
if __name__=='__main__':
import argparse
parser=argparse.ArgumentParser();parser.add_argument('index');args=parser.parse_args()
dataset=TrainingDataset(args.index)
for index in (0,len(dataset)//2,len(dataset)-1):
row=dataset[index];print(row['task'],row['frame'],{name:r['state'].shape for name,r in row['robots'].items()},len(row['images']))
print(len(dataset.rows),'accepted episodes',len(dataset),'aligned transitions');dataset.close()
|