eval run 20260904_194409 -> eval_runs/20260904_194409
camera map: {'gripper': 'fanuc_01', 'cam0': 'fanuc_02', 'cam1': 'fanuc_03', 'cam2': 'fanuc_04'}
task: Pick up the dice and place it on the empty white block.
cell_pause: a collection is running - waiting for it to finish the episode, park and release the stream
cell_pause: collector paused (collect_full, session s20260904_131844) - the arm is ours
policy: connected at /opt/policy/ckpt/imle_stream_orig/checkpoints/080000/pretrained_model
  action contract: joint_abs (absolute joint degrees), 7 wide: J1, J2, J3, J4, J5, J6, grip
  running in-process on cuda, views cam0, gripper
task: not sent - this checkpoint is not task-conditioned (no environment_state feature); 'Pick up the dice and place it on the empty white block.' is recorded for provenance only
  state layout: j7, 7 wide (joint degrees, gripper 1.0 = OPEN)
  observation frames: RAW shared memory ['fanuc_01', 'fanuc_02'] - the channel the corpus was recorded on
  frames: raw (undistort none)
  undistort: 'none' matches training dataset dice_stream_j7
  [WARNING] camera_map keys ['cam1', 'cam2'] are NOT policy inputs (this policy takes ['gripper', 'cam0']) - they are recorded but never sent
  observation frames: raw pixels, no codec round trip
observation ok: sent {'cam0': (480, 640, 3), 'gripper': (480, 640, 3)}
  cameras delivered: {'cam0': (480, 640, 3), 'gripper': (480, 640, 3)}
  jpeg quality: 80 (the corpus is lossless renders; raise it with the service's --quality if you are chasing a domain gap)
  start-height gate: OFF (start_z_max_mm is null) - the failed runs began ~200 mm above anything the corpus visits and nothing here would catch that
  state-range gate: ARMED, 6 channels; the arm is currently INSIDE the corpus range
  memory latch: n/a - this policy has no server-side first-frame view
parked at home before starting
controller alarm log: md:/errall.ls on 192.168.1.100, diffed per episode by controller time
=== episode 1/1 ===
  1 of 2 blocks bare -> dice is on a block
    dice trim h0 (z -91.5): off 2.20 mm
    dice trim h1 (z -91.5): off 0.54 mm
    dice view check: 2.57 mm off at grasp depth (over the 1.8 gate - the taught references disagree by a position-dependent amount; grasping on the height convergence)
     lifting to z -20 before touching the wrist (rotating at z -127 spins the part against its neighbours)
     lifting 55 mm clear of the approach zone first
    block align 0: block at (256,179) area 6423 yaw +3.5 deg -> off 13.70 mm
    block align 1: block at (267,235) area 7642 yaw +1.3 deg -> off 1.09 mm
    block align 2: block at (271,237) area 7514 yaw +1.8 deg -> off 0.24 mm
     lifting to z -20 before touching the wrist (rotating at z -100 spins the part against its neighbours)
     lifting 55 mm clear of the approach zone first
  dice (341.8, 435.2)  bare block (235.8, 528.2) [+/-0.24 mm]
  oracle: dice [np.float64(341.85), np.float64(435.2)] -> bare block [np.float64(235.84), np.float64(528.21)]
  running the policy (timeout 600 s)
  recording 4 feeds at 15 fps
  start pose: tool 23.06 deg off vertical
  transport: Stream Motion, whole chunks at 125 Hz
  observations: pairs 33 ms apart (this checkpoint takes 2 steps)
  recorded 639 frames per feed (0 dropped)
     lifting to z -20 before touching the wrist (rotating at z -96 spins the part against its neighbours)
  policy: budget after 1000 steps / 33.8 s
  controller alarms during episode: none
  1 of 2 blocks bare -> dice is on a block
    dice trim h0 (z -91.5): off 3.12 mm
    dice trim h1 (z -91.5): off 0.43 mm
    dice view check: 2.69 mm off at grasp depth (over the 1.8 gate - the taught references disagree by a position-dependent amount; grasping on the height convergence)
     lifting to z -20 before touching the wrist (rotating at z -127 spins the part against its neighbours)
     lifting 55 mm clear of the approach zone first
    block align 0: block at (256,170) area 6574 yaw +3.4 deg -> off 15.61 mm
    block align 1: block at (267,235) area 7632 yaw +2.4 deg -> off 1.04 mm
    block align 2: block at (271,236) area 7538 yaw +1.6 deg -> off 0.30 mm
     lifting to z -20 before touching the wrist (rotating at z -100 spins the part against its neighbours)
     lifting 55 mm clear of the approach zone first
  dice (340.8, 435.2)  bare block (235.6, 528.3) [+/-0.30 mm]
  result: on_block=True, 140.3 mm from the target block -> failure
    APPROACH_NO_GRASP | stage 3/6 | progress 0.39 | cos_align +0.61 | path 41981 mm | grip max 0.00
    (resampling: 447, 365 is outside the caller's placement hull)
    spot (239, 370), 121 mm clear of everything (1 blacklisted zone honoured)
  re-randomising the bare block -> (238.9, 370.3)
    block pre-grasp: square (-1.5 deg off)
    block align 0: block at (272,236) area 7490 yaw +1.7 deg -> off 0.06 mm
  far-reach descent: crawling at 25 mm/s
    block depth check: off 0.21 mm
  block grasped at (235.5, 528.3)
    block squaring: square (-1.3 deg off)
parked at home

=== 0/1 succeeded ===
  success 0%  (95% Wilson 0-79%, width 79 pts)
  posterior  median 29%  (95% Beta 1-84%)
  scoring success_at_end (the dice must still be on the block)

  reached at least stage k:
    stage 0   1/1  #
    stage 1   1/1  #
    stage 2   1/1  #
    stage 3   1/1  #
    stage 4   0/1  
    stage 5   0/1  
    stage 6   0/1  

  failure classes:
    APPROACH_NO_GRASP   1  (air_grasp x1)

  progress   0.390   (95% bootstrap 0.390-0.390; 1.0 = reached the dice)
  cos_align  +0.610  (+1 straight at the dice, 0 uncorrelated, -1 away)
  toward     86% of steps
  path       41981 mm/episode
  d_min      8 mm   clamp rate nan   gripper max 0.00
  timing     median reach 13 s, grasp -- s (among episodes that got there)
  jitter     follow_max median 3.14 deg, reversals median 0, stretch max 2.5
=== 0/1 succeeded (0%) ===
uploading eval_runs/20260904_194409 to the Hub ...